diff --git a/feed_weekly.xml b/feed_weekly.xml index 6637531..d6fcae2 100644 --- a/feed_weekly.xml +++ b/feed_weekly.xml @@ -7,741 +7,741 @@ http://www.rssboard.org/rss-specification python-feedgen en - Mon, 09 Mar 2026 00:15:39 +0000 + Mon, 16 Mar 2026 00:19:28 +0000 - Spilled Energy in Large Language Models - https://arxiv.org/abs/2602.18671 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.18671.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Adrian Robert Minut, Hazem Dewidar, Iacopo Masi</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> We reinterpret the final Large Language Model (LLM) softmax classifier as an Energy-Based Model (EBM), decomposing the sequence-to-sequence probability chain into multiple interacting EBMs at inference. This principled approach allows us to track "energy spills" during decoding, which we empirically show correlate with factual errors, biases, and failures. Similar to Orgad et al. (2025), our method localizes the exact answer token and subsequently tests for hallucinations. Crucially, however, we achieve this without requiring trained probe classifiers or activation ablations. Instead, we introduce two completely training-free metrics derived directly from output logits: spilled energy, which captures the discrepancy between energy values across consecutive generation steps that should theoretically match, and marginalized energy, which is measurable at a single step. Evaluated on nine benchmarks across state-of-the-art LLMs (including LLaMA, Mistral, and Gemma) and on synthetic algebraic operations (Qwen3), our approach demonstrates robust, competitive hallucination detection and cross-task generalization. Notably, these results hold for both pretrained and instruction-tuned variants without introducing any training overhead. Code available at: github.com/OmnAI-Lab/spilled-energy</p> - https://arxiv.org/abs/2602.18671 - Sat, 21 Feb 2026 00:38:47 +0000 + WildActor: Unconstrained Identity-Preserving Video Generation + https://arxiv.org/abs/2603.00586 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00586.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qin Guo, Tianyu Yang, Xuanhua He, Fei Shen, Yong Zhang, Zhuoliang Kang, Xiaoming Wei, Dan Xu</p><p><b>Upvotes:</b> 34</p><p><b>Summary:</b> Production-ready human video generation requires digital actors to maintain strictly consistent full-body identities across dynamic shots, viewpoints and motions, a setting that remains challenging for existing methods. Prior methods often suffer from face-centric behavior that neglects body-level consistency, or produce copy-paste artifacts where subjects appear rigid due to pose locking. We present Actor-18M, a large-scale human video dataset designed to capture identity consistency under unconstrained viewpoints and environments. Actor-18M comprises 1.6M videos with 18M corresponding human images, covering both arbitrary views and canonical three-view representations. Leveraging Actor-18M, we propose WildActor, a framework for any-view conditioned human video generation. We introduce an Asymmetric Identity-Preserving Attention mechanism coupled with a Viewpoint-Adaptive Monte Carlo Sampling strategy that iteratively re-weights reference conditions by marginal utility for balanced manifold coverage. Evaluated on the proposed Actor-Bench, WildActor consistently preserves body identity under diverse shot compositions, large viewpoint transitions, and substantial motions, surpassing existing methods in these challenging settings.</p> + https://arxiv.org/abs/2603.00586 + Sat, 28 Feb 2026 10:34:39 +0000 - Classroom Final Exam: An Instructor-Tested Reasoning Benchmark - https://arxiv.org/abs/2602.19517 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.19517.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chongyang Gao, Diji Yang, Shuyan Zhou, Xichen Yan, Luchuan Song, Shuo Li, Kezhen Chen</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> We introduce (Classroom Final Exam), a multimodal benchmark for evaluating the reasoning capabilities of large language models across more than 20 STEM domains. is curated from repeatedly used, authentic university homework and exam problems, together with reference solutions provided by course instructors. presents a significant challenge even for frontier models: the newly released Gemini-3.1-pro-preview achieves an overall accuracy of 59.69\%, while the second-best model, Gemini-3-flash-preview, reaches 55.46\%, leaving considerable room for improvement. Beyond leaderboard results, we perform a diagnostic analysis by decomposing reference solutions into reasoning flows. We find that although frontier models can often answer intermediate sub-questions correctly, they struggle to reliably derive and maintain correct intermediate states throughout multi-step solutions. We further observe that model-generated solutions typically have more reasoning steps than those provided by the instructor, indicating suboptimal step efficiency and a higher risk of error accumulation. The data and code are available at https://github.com/Analogy-AI/CFE_Bench.</p> - https://arxiv.org/abs/2602.19517 - Mon, 23 Feb 2026 05:17:41 +0000 + π-StepNFT: Wider Space Needs Finer Steps in Online RL for Flow-based VLAs + https://arxiv.org/abs/2603.02083 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02083.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Siting Wang, Xiaofeng Wang, Zheng Zhu, Minnan Pei, Xinyu Cui, Cheng Deng, Jian Zhao, Guan Huang, Haifeng Zhang, Jun Wang</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Flow-based vision-language-action (VLA) models excel in embodied control but suffer from intractable likelihoods during multi-step sampling, hindering online reinforcement learning. We propose \textit{boldsymbolπ-StepNFT} (Step-wise Negative-aware Fine-Tuning), a critic-and-likelihood-free framework that requires only a single forward pass per optimization step and eliminates auxiliary value networks. We identify that wider exploration spaces necessitate finer-grained, step-wise guidance for alignment. Empirically, π-StepNFT unlocks latent potential on LIBERO with competitive few-shot robustness. Moreover, it achieves superior generalization on ManiSkill, outperforming value-based baselines in OOD scenarios by preventing overfitting to multimodal features. This property offers a scalable solution promising for complex real-world applications.</p> + https://arxiv.org/abs/2603.02083 + Mon, 02 Mar 2026 17:04:49 +0000 - From Scale to Speed: Adaptive Test-Time Scaling for Image Editing - https://arxiv.org/abs/2603.00141 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00141.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiangyan Qu, Zhenlong Yuan, Jing Tang, Rui Chen, Datao Tang, Meng Yu, Lei Sun, Yancheng Bai, Xiangxiang Chu, Gaopeng Gou, Gang Xiong, Yujun Cai</p><p><b>Upvotes:</b> 130</p><p><b>Summary:</b> Image Chain-of-Thought (Image-CoT) is a test-time scaling paradigm that improves image generation by extending inference time. Most Image-CoT methods focus on text-to-image (T2I) generation. Unlike T2I generation, image editing is goal-directed: the solution space is constrained by the source image and instruction. This mismatch causes three challenges when applying Image-CoT to editing: inefficient resource allocation with fixed sampling budgets, unreliable early-stage verification using general MLLM scores, and redundant edited results from large-scale sampling. To address this, we propose ADaptive Edit-CoT (ADE-CoT), an on-demand test-time scaling framework to enhance editing efficiency and performance. It incorporates three key strategies: (1) a difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty; (2) edit-specific verification in early pruning that uses region localization and caption consistency to select promising candidates; and (3) depth-first opportunistic stopping, guided by an instance-specific verifier, that terminates when intent-aligned results are found. Extensive experiments on three SOTA editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show that ADE-CoT achieves superior performance-efficiency trade-offs. With comparable sampling budgets, ADE-CoT obtains better performance with more than 2x speedup over Best-of-N.</p> - https://arxiv.org/abs/2603.00141 - Tue, 24 Feb 2026 09:13:16 +0000 + Geometry-Guided Reinforcement Learning for Multi-view Consistent 3D Scene Editing + https://arxiv.org/abs/2603.03143 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03143.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiyuan Wang, Chunyu Lin, Lei Sun, Zhi Cao, Yuyang Yin, Lang Nie, Zhenlong Yuan, Xiangxiang Chu, Yunchao Wei, Kang Liao, Guosheng Lin</p><p><b>Upvotes:</b> 136</p><p><b>Summary:</b> Leveraging the priors of 2D diffusion models for 3D editing has emerged as a promising paradigm. However, maintaining multi-view consistency in edited results remains challenging, and the extreme scarcity of 3D-consistent editing paired data renders supervised fine-tuning (SFT), the most effective training strategy for editing tasks, infeasible. In this paper, we observe that, while generating multi-view consistent 3D content is highly challenging, verifying 3D consistency is tractable, naturally positioning reinforcement learning (RL) as a feasible solution. Motivated by this, we propose RL3DEdit, a single-pass framework driven by RL optimization with novel rewards derived from the 3D foundation model, VGGT. Specifically, we leverage VGGT's robust priors learned from massive real-world data, feed the edited images, and utilize the output confidence maps and pose estimation errors as reward signals, effectively anchoring the 2D editing priors onto a 3D-consistent manifold via RL. Extensive experiments demonstrate that RL3DEdit achieves stable multi-view consistency and outperforms state-of-the-art methods in editing quality with high efficiency. To promote the development of 3D editing, we will release the code and model.</p> + https://arxiv.org/abs/2603.03143 + Tue, 03 Mar 2026 16:31:10 +0000 - LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding - https://arxiv.org/abs/2602.20913 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.20913.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jihao Qiu, Lingxi Xie, Xinyue Huo, Qi Tian, Qixiang Ye</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> This paper addresses the critical and underexplored challenge of long video understanding with low computational budgets. We propose LongVideo-R1, an active, reasoning-equipped multimodal large language model (MLLM) agent designed for efficient video context navigation, avoiding the redundancy of exhaustive search. At the core of LongVideo-R1 lies a reasoning module that leverages high-level visual cues to infer the most informative video clip for subsequent processing. During inference, the agent initiates traversal from top-level visual summaries and iteratively refines its focus, immediately halting the exploration process upon acquiring sufficient knowledge to answer the query. To facilitate training, we first extract hierarchical video captions from CGBench, a video corpus with grounding annotations, and guide GPT-5 to generate 33K high-quality chain-of-thought-with-tool trajectories. The LongVideo-R1 agent is fine-tuned upon the Qwen-3-8B model through a two-stage paradigm: supervised fine-tuning (SFT) followed by reinforcement learning (RL), where RL employs a specifically designed reward function to maximize selective and efficient clip navigation. Experiments on multiple long video benchmarks validate the effectiveness of name, which enjoys superior tradeoff between QA accuracy and efficiency. All curated data and source code are provided in the supplementary material and will be made publicly available. Code and data are available at: https://github.com/qiujihao19/LongVideo-R1</p> - https://arxiv.org/abs/2602.20913 - Tue, 24 Feb 2026 13:49:47 +0000 + LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory + https://arxiv.org/abs/2603.03269 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03269.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junyi Zhang, Charles Herrmann, Junhwa Hur, Chen Sun, Ming-Hsuan Yang, Forrester Cole, Trevor Darrell, Deqing Sun</p><p><b>Upvotes:</b> 53</p><p><b>Summary:</b> Feedforward geometric foundation models achieve strong short-window reconstruction, yet scaling them to minutes-long videos is bottlenecked by quadratic attention complexity or limited effective memory in recurrent designs. We present LoGeR (Long-context Geometric Reconstruction), a novel architecture that scales dense 3D reconstruction to extremely long sequences without post-optimization. LoGeR processes video streams in chunks, leveraging strong bidirectional priors for high-fidelity intra-chunk reasoning. To manage the critical challenge of coherence across chunk boundaries, we propose a learning-based hybrid memory module. This dual-component system combines a parametric Test-Time Training (TTT) memory to anchor the global coordinate frame and prevent scale drift, alongside a non-parametric Sliding Window Attention (SWA) mechanism to preserve uncompressed context for high-precision adjacent alignment. Remarkably, this memory architecture enables LoGeR to be trained on sequences of 128 frames, and generalize up to thousands of frames during inference. Evaluated across standard benchmarks and a newly repurposed VBR dataset with sequences of up to 19k frames, LoGeR substantially outperforms prior state-of-the-art feedforward methods--reducing ATE on KITTI by over 74%--and achieves robust, globally consistent reconstruction over unprecedented horizons.</p> + https://arxiv.org/abs/2603.03269 + Tue, 03 Mar 2026 18:55:37 +0000 - Shared Nature, Unique Nurture: PRISM for Pluralistic Reasoning via In-context Structure Modeling - https://arxiv.org/abs/2602.21317 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.21317.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Guancheng Tu, Shiyang Zhang, Tianyu Zhang, Yi Zhang, Diji Yang</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Large Language Models (LLMs) are converging towards a singular Artificial Hivemind, where shared Nature (pre-training priors) result in a profound collapse of distributional diversity, limiting the distinct perspectives necessary for creative exploration and scientific discovery. To address this, we propose to equip models with inference-time Nurture (individualized epistemic trajectories) using Epistemic Evolution paradigm, progressing through explore, internalize, and express. We instantiate this via PRISM (Pluralistic Reasoning via In-context Structure Modeling), a model-agnostic system that augments LLM with dynamic On-the-fly Epistemic Graphs. On three creativity benchmarks, PRISM achieves state-of-the-art novelty and significantly expands distributional diversity. Moreover, we evaluate the real-world utility via a challenging rare-disease diagnosis benchmark. Results demonstrate that PRISM successfully uncovers correct long-tail diagnoses that standard LLM miss, confirming that its divergence stems from meaningful exploration rather than incoherent noise. Overall, this work establishes a new paradigm for Pluralistic AI, moving beyond monolithic consensus toward a diverse ecosystem of unique cognitive individuals capable of collective, multi-perspective discovery.</p> - https://arxiv.org/abs/2602.21317 - Tue, 24 Feb 2026 19:38:31 +0000 + From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning + https://arxiv.org/abs/2603.03825 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03825.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ruilin Luo, Chufan Shi, Yizhen Zhang, Cheng Yang, Songtao Jiang, Tongkun Guan, Ruizhe Chen, Ruihang Chu, Peng Wang, Mingkun Yang, Yujiu Yang, Junyang Lin, Zhibo Yang</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> The cold-start initialization stage plays a pivotal role in training Multimodal Large Reasoning Models (MLRMs), yet its mechanisms remain insufficiently understood. To analyze this stage, we introduce the Visual Attention Score (VAS), an attention-based metric that quantifies how much a model attends to visual tokens. We find that reasoning performance is strongly correlated with VAS (r=0.9616): models with higher VAS achieve substantially stronger multimodal reasoning. Surprisingly, multimodal cold-start fails to elevate VAS, resulting in attention distributions close to the base model, whereas text-only cold-start leads to a clear increase. We term this counter-intuitive phenomenon Lazy Attention Localization. To validate its causal role, we design training-free interventions that directly modulate attention allocation during inference, performance gains of 1-2% without any retraining. Building on these insights, we further propose Attention-Guided Visual Anchoring and Reflection (AVAR), a comprehensive cold-start framework that integrates visual-anchored data synthesis, attention-guided objectives, and visual-anchored reward shaping. Applied to Qwen2.5-VL-7B, AVAR achieves an average gain of 7.0% across 7 multimodal reasoning benchmarks. Ablation studies further confirm that each component of AVAR contributes step-wise to the overall gains. The code, data, and models are available at https://github.com/lrlbbzl/Qwen-AVAR.</p> + https://arxiv.org/abs/2603.03825 + Wed, 04 Mar 2026 08:22:27 +0000 - Tool-R0: Self-Evolving LLM Agents for Tool-Learning from Zero Data - https://arxiv.org/abs/2602.21320 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.21320.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Emre Can Acikgoz, Cheng Qian, Jonas Hübotter, Heng Ji, Dilek Hakkani-Tür, Gokhan Tur</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Large language models (LLMs) are becoming the foundation for autonomous agents that can use tools to solve complex tasks. Reinforcement learning (RL) has emerged as a common approach for injecting such agentic capabilities, but typically under tightly controlled training setups. It often depends on carefully constructed task-solution pairs and substantial human supervision, which creates a fundamental obstacle to open-ended self-evolution toward superintelligent systems. In this paper, we propose Tool-R0 framework for training general purpose tool-calling agents from scratch with self-play RL, under a zero-data assumption. Initialized from the same base LLM, Tool-R0 co-evolves a Generator and a Solver with complementary rewards: one proposes targeted challenging tasks at the other's competence frontier and the other learns to solve them with real-world tool calls. This creates a self-evolving cycle that requires no pre-existing tasks or datasets. Evaluation on different tool-use benchmarks show that Tool-R0 yields 92.5 relative improvement over the base model and surpasses fully supervised tool-calling baselines under the same setting. Our work further provides empirical insights into self-play LLM agents by analyzing co-evolution, curriculum dynamics, and scaling behavior.</p> - https://arxiv.org/abs/2602.21320 - Tue, 24 Feb 2026 19:41:18 +0000 + Believe Your Model: Distribution-Guided Confidence Calibration + https://arxiv.org/abs/2603.03872 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03872.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xizhong Yang, Haotian Zhang, Huiming Wang, Mofei Song</p><p><b>Upvotes:</b> 38</p><p><b>Summary:</b> Large Reasoning Models have demonstrated remarkable performance with the advancement of test-time scaling techniques, which enhances prediction accuracy by generating multiple candidate responses and selecting the most reliable answer. While prior work has analyzed that internal model signals like confidence scores can partly indicate response correctness and exhibit a distributional correlation with accuracy, such distributional information has not been fully utilized to guide answer selection. Motivated by this, we propose DistriVoting, which incorporates distributional priors as another signal alongside confidence during voting. Specifically, our method (1) first decomposes the mixed confidence distribution into positive and negative components using Gaussian Mixture Models, (2) then applies a reject filter based on positive/negative samples from them to mitigate overlap between the two distributions. Besides, to further alleviate the overlap from the perspective of distribution itself, we propose SelfStepConf, which uses step-level confidence to dynamically adjust inference process, increasing the separation between the two distributions to improve the reliability of confidences in voting. Experiments across 16 models and 5 benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches.</p> + https://arxiv.org/abs/2603.03872 + Wed, 04 Mar 2026 09:25:36 +0000 - How to Take a Memorable Picture? Empowering Users with Actionable Feedback - https://arxiv.org/abs/2602.21877 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.21877.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Francesco Laiti, Davide Talon, Jacopo Staiano, Elisa Ricci</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Image memorability, i.e., how likely an image is to be remembered, has traditionally been studied in computer vision either as a passive prediction task, with models regressing a scalar score, or with generative methods altering the visual input to boost the image likelihood of being remembered. Yet, none of these paradigms supports users at capture time, when the crucial question is how to improve a photo memorability. We introduce the task of Memorability Feedback (MemFeed), where an automated model should provide actionable, human-interpretable guidance to users with the goal to enhance an image future recall. We also present MemCoach, the first approach designed to provide concrete suggestions in natural language for memorability improvement (e.g., "emphasize facial expression," "bring the subject forward"). Our method, based on Multimodal Large Language Models (MLLMs), is training-free and employs a teacher-student steering strategy, aligning the model internal activations toward more memorable patterns learned from a teacher model progressing along least-to-most memorable samples. To enable systematic evaluation on this novel task, we further introduce MemBench, a new benchmark featuring sequence-aligned photoshoots with annotated memorability scores. Our experiments, considering multiple MLLMs, demonstrate the effectiveness of MemCoach, showing consistently improved performance over several zero-shot models. The results indicate that memorability can not only be predicted but also taught and instructed, shifting the focus from mere prediction to actionable feedback for human creators.</p> - https://arxiv.org/abs/2602.21877 - Wed, 25 Feb 2026 13:02:35 +0000 + Bootstrapping Exploration with Group-Level Natural Language Feedback in Reinforcement Learning + https://arxiv.org/abs/2603.04597 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04597.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lei Huang, Xiang Cheng, Chenxiao Zhao, Guobin Shen, Junjie Yang, Xiaocheng Feng, Yuxuan Gu, Xing Yu, Bing Qin</p><p><b>Upvotes:</b> 156</p><p><b>Summary:</b> Large language models (LLMs) typically receive diverse natural language (NL) feedback through interaction with the environment. However, current reinforcement learning (RL) algorithms rely solely on scalar rewards, leaving the rich information in NL feedback underutilized and leading to inefficient exploration. In this work, we propose GOLF, an RL framework that explicitly exploits group-level language feedback to guide targeted exploration through actionable refinements. GOLF aggregates two complementary feedback sources: (i) external critiques that pinpoint errors or propose targeted fixes, and (ii) intra-group attempts that supply alternative partial ideas and diverse failure patterns. These group-level feedbacks are aggregated to produce high-quality refinements, which are adaptively injected into training as off-policy scaffolds to provide targeted guidance in sparse-reward regions. Meanwhile, GOLF jointly optimizes generation and refinement within a unified RL loop, creating a virtuous cycle that continuously improves both capabilities. Experiments on both verifiable and non-verifiable benchmarks show that GOLF achieves superior performance and exploration efficiency, achieving 2.2times improvements in sample efficiency compared to RL methods trained solely on scalar rewards. Code is available at https://github.com/LuckyyySTA/GOLF.</p> + https://arxiv.org/abs/2603.04597 + Wed, 04 Mar 2026 20:53:17 +0000 - Recovered in Translation: Efficient Pipeline for Automated Translation of Benchmarks and Datasets - https://arxiv.org/abs/2602.22207 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.22207.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hanna Yukhymenko, Anton Alexandrov, Martin Vechev</p><p><b>Upvotes:</b> 39</p><p><b>Summary:</b> The reliability of multilingual Large Language Model (LLM) evaluation is currently compromised by the inconsistent quality of translated benchmarks. Existing resources often suffer from semantic drift and context loss, which can lead to misleading performance metrics. In this work, we present a fully automated framework designed to address these challenges by enabling scalable, high-quality translation of datasets and benchmarks. We demonstrate that adapting test-time compute scaling strategies, specifically Universal Self-Improvement (USI) and our proposed multi-round ranking method, T-RANK, allows for significantly higher quality outputs compared to traditional pipelines. Our framework ensures that benchmarks preserve their original task structure and linguistic nuances during localization. We apply this approach to translate popular benchmarks and datasets into eight Eastern and Southern European languages (Ukrainian, Bulgarian, Slovak, Romanian, Lithuanian, Estonian, Turkish, Greek). Evaluations using both reference-based metrics and LLM-as-a-judge show that our translations surpass existing resources, resulting in more accurate downstream model assessment. We release both the framework and the improved benchmarks to facilitate robust and reproducible multilingual AI development.</p> - https://arxiv.org/abs/2602.22207 - Wed, 25 Feb 2026 18:58:25 +0000 + RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies + https://arxiv.org/abs/2603.04639 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04639.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yinpei Dai, Hongze Fu, Jayjun Lee, Yuejiang Liu, Haoran Zhang, Jianing Yang, Chelsea Finn, Nima Fazeli, Joyce Chai</p><p><b>Upvotes:</b> 23</p><p><b>Summary:</b> Memory is critical for long-horizon and history-dependent robotic manipulation. Such tasks often involve counting repeated actions or manipulating objects that become temporarily occluded. Recent vision-language-action (VLA) models have begun to incorporate memory mechanisms; however, their evaluations remain confined to narrow, non-standardized settings. This limits their systematic understanding, comparison, and progress measurement. To address these challenges, we introduce RoboMME: a large-scale standardized benchmark for evaluating and advancing VLA models in long-horizon, history-dependent scenarios. Our benchmark comprises 16 manipulation tasks constructed under a carefully designed taxonomy that evaluates temporal, spatial, object, and procedural memory. We further develop a suite of 14 memory-augmented VLA variants built on the π0.5 backbone to systematically explore different memory representations across multiple integration strategies. Experimental results show that the effectiveness of memory representations is highly task-dependent, with each design offering distinct advantages and limitations across different tasks. Videos and code can be found at our website https://robomme.github.io.</p> + https://arxiv.org/abs/2603.04639 + Wed, 04 Mar 2026 21:59:32 +0000 - dLLM: Simple Diffusion Language Modeling - https://arxiv.org/abs/2602.22661 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.22661.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhanhui Zhou, Lingjie Chen, Hanghang Tong, Dawn Song</p><p><b>Upvotes:</b> 121</p><p><b>Summary:</b> Although diffusion language models (DLMs) are evolving quickly, many recent models converge on a set of shared components. These components, however, are distributed across ad-hoc research codebases or lack transparent implementations, making them difficult to reproduce or extend. As the field accelerates, there is a clear need for a unified framework that standardizes these common components while remaining flexible enough to support new methods and architectures. To address this gap, we introduce dLLM, an open-source framework that unifies the core components of diffusion language modeling -- training, inference, and evaluation -- and makes them easy to customize for new designs. With dLLM, users can reproduce, finetune, deploy, and evaluate open-source large DLMs such as LLaDA and Dream through a standardized pipeline. The framework also provides minimal, reproducible recipes for building small DLMs from scratch with accessible compute, including converting any BERT-style encoder or autoregressive LM into a DLM. We also release the checkpoints of these small DLMs to make DLMs more accessible and accelerate future research.</p> - https://arxiv.org/abs/2602.22661 - Thu, 26 Feb 2026 06:26:02 +0000 + HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel + https://arxiv.org/abs/2603.04750 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04750.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> The Viet Bui, Wenjun Li, Yong Liu</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Sequential LLM agents fail on long-horizon planning with hard constraints like budgets and diversity requirements. As planning progresses and context grows, these agents drift from global constraints. We propose HiMAP-Travel, a hierarchical multi-agent framework that splits planning into strategic coordination and parallel day-level execution. A Coordinator allocates resources across days, while Day Executors plan independently in parallel. Three key mechanisms enable this: a transactional monitor enforcing budget and uniqueness constraints across parallel agents, a bargaining protocol allowing agents to reject infeasible sub-goals and trigger re-planning, and a single policy trained with GRPO that powers all agents through role conditioning. On TravelPlanner, HiMAP-Travel with Qwen3-8B achieves 52.78% validation and 52.65% test Final Pass Rate (FPR). In a controlled comparison with identical model, training, and tools, it outperforms the sequential DeepTravel baseline by +8.67~pp. It also surpasses ATLAS by +17.65~pp and MTP by +10.0~pp. On FlexTravelBench multi-turn scenarios, it achieves 44.34% (2-turn) and 37.42% (3-turn) FPR while reducing latency 2.5x through parallelization.</p> + https://arxiv.org/abs/2603.04750 + Thu, 05 Mar 2026 02:55:53 +0000 - Towards Simulating Social Media Users with LLMs: Evaluating the Operational Validity of Conditioned Comment Prediction - https://arxiv.org/abs/2602.22752 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.22752.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Nils Schwager, Simon Münker, Alistair Plum, Achim Rettinger</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> The transition of Large Language Models (LLMs) from exploratory tools to active "silicon subjects" in social science lacks extensive validation of operational validity. This study introduces Conditioned Comment Prediction (CCP), a task in which a model predicts how a user would comment on a given stimulus by comparing generated outputs with authentic digital traces. This framework enables a rigorous evaluation of current LLM capabilities with respect to the simulation of social media user behavior. We evaluated open-weight 8B models (Llama3.1, Qwen3, Ministral) in English, German, and Luxembourgish language scenarios. By systematically comparing prompting strategies (explicit vs. implicit) and the impact of Supervised Fine-Tuning (SFT), we identify a critical form vs. content decoupling in low-resource settings: while SFT aligns the surface structure of the text output (length and syntax), it degrades semantic grounding. Furthermore, we demonstrate that explicit conditioning (generated biographies) becomes redundant under fine-tuning, as models successfully perform latent inference directly from behavioral histories. Our findings challenge current "naive prompting" paradigms and offer operational guidelines prioritizing authentic behavioral traces over descriptive personas for high-fidelity simulation.</p> - https://arxiv.org/abs/2602.22752 - Thu, 26 Feb 2026 08:40:21 +0000 + BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning + https://arxiv.org/abs/2603.04918 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04918.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuan Li, Bo Wang, Yufei Gao, Yuqian Yao, Xinyuan Wang, Zhangyue Yin, Xipeng Qiu</p><p><b>Upvotes:</b> 54</p><p><b>Summary:</b> Proximal constraints are fundamental to the stability of the Large Language Model reinforcement learning. While the canonical clipping mechanism in PPO serves as an efficient surrogate for trust regions, we identify a critical bottleneck: fixed bounds strictly constrain the upward update margin of low-probability actions, disproportionately suppressing high-advantage tail strategies and inducing rapid entropy collapse. To address this, we introduce Band-constrained Policy Optimization (BandPO). BandPO replaces canonical clipping with Band, a unified theoretical operator that projects trust regions defined by f-divergences into dynamic, probability-aware clipping intervals. Theoretical analysis confirms that Band effectively resolves this exploration bottleneck. We formulate this mapping as a convex optimization problem, guaranteeing a globally optimal numerical solution while deriving closed-form solutions for specific divergences. Extensive experiments across diverse models and datasets demonstrate that BandPO consistently outperforms canonical clipping and Clip-Higher, while robustly mitigating entropy collapse.</p> + https://arxiv.org/abs/2603.04918 + Thu, 05 Mar 2026 08:03:05 +0000 - SkillNet: Create, Evaluate, and Connect AI Skills - https://arxiv.org/abs/2603.04448 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04448.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuan Liang, Ruobin Zhong, Haoming Xu, Chen Jiang, Yi Zhong, Runnan Fang, Jia-Chen Gu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, Xin Xu, Tongtong Wu, Kun Wang, Yang Liu, Zhen Bi, Jungang Lou, Yuchen Eleanor Jiang, Hangcheng Zhu, Gang Yu, Haiwen Hong, Longtao Huang, Hui Xue, Chenxi Wang, Yijun Wang, Zifei Shan, Xi Chen, Zhaopeng Tu, Feiyu Xiong, Xin Xie, Peng Zhang, Zhengke Gui, Lei Liang, Jun Zhou, Chiyu Wu, Jin Shang, Yu Gong, Junyu Lin, Changliang Xu, Hongjie Deng, Wen Zhang, Keyan Ding, Qiang Zhang, Fei Huang, Ningyu Zhang, Jeff Z. Pan, Guilin Qi, Haofen Wang, Huajun Chen</p><p><b>Upvotes:</b> 64</p><p><b>Summary:</b> Current AI agents can flexibly invoke tools and execute complex tasks, yet their long-term advancement is hindered by the lack of systematic accumulation and transfer of skills. Without a unified mechanism for skill consolidation, agents frequently ``reinvent the wheel'', rediscovering solutions in isolated contexts without leveraging prior strategies. To overcome this limitation, we introduce SkillNet, an open infrastructure designed to create, evaluate, and organize AI skills at scale. SkillNet structures skills within a unified ontology that supports creating skills from heterogeneous sources, establishing rich relational connections, and performing multi-dimensional evaluation across Safety, Completeness, Executability, Maintainability, and Cost-awareness. Our infrastructure integrates a repository of over 200,000 skills, an interactive platform, and a versatile Python toolkit. Experimental evaluations on ALFWorld, WebShop, and ScienceWorld demonstrate that SkillNet significantly enhances agent performance, improving average rewards by 40% and reducing execution steps by 30% across multiple backbone models. By formalizing skills as evolving, composable assets, SkillNet provides a robust foundation for agents to move from transient experience to durable mastery.</p> - https://arxiv.org/abs/2603.04448 - Thu, 26 Feb 2026 14:24:02 +0000 + Sparse-BitNet: 1.58-bit LLMs are Naturally Friendly to Semi-Structured Sparsity + https://arxiv.org/abs/2603.05168 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05168.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Di Zhang, Xun Wu, Shaohan Huang, Yudong Wang, Hanyong Shao, Yingbo Hao, Zewen Chi, Li Dong, Ting Song, Yan Xia, Zhifang Sui, Furu Wei</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Semi-structured N:M sparsity and low-bit quantization (e.g., 1.58-bit BitNet) are two promising approaches for improving the efficiency of large language models (LLMs), yet they have largely been studied in isolation. In this work, we investigate their interaction and show that 1.58-bit BitNet is naturally more compatible with N:M sparsity than full-precision models. To study this effect, we propose Sparse-BitNet, a unified framework that jointly applies 1.58-bit quantization and dynamic N:M sparsification while ensuring stable training for the first time. Across multiple model scales and training regimes (sparse pretraining and dense-to-sparse schedules), 1.58-bit BitNet consistently exhibits smaller performance degradation than full-precision baselines at the same sparsity levels and can tolerate higher structured sparsity before accuracy collapse. Moreover, using our custom sparse tensor core, Sparse-BitNet achieves substantial speedups in both training and inference, reaching up to 1.30X. These results highlight that combining extremely low-bit quantization with semi-structured N:M sparsity is a promising direction for efficient LLMs. Code available at https://github.com/AAzdi/Sparse-BitNet</p> + https://arxiv.org/abs/2603.05168 + Thu, 05 Mar 2026 13:37:50 +0000 - AgentVista: Evaluating Multimodal Agents in Ultra-Challenging Realistic Visual Scenarios - https://arxiv.org/abs/2602.23166 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23166.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhaochen Su, Jincheng Gao, Hangyu Guo, Zhenhua Liu, Lueyang Zhang, Xinyu Geng, Shijue Huang, Peng Xia, Guanyu Jiang, Cheng Wang, Yue Zhang, Yi R. Fung, Junxian He</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> Real-world multimodal agents solve multi-step workflows grounded in visual evidence. For example, an agent can troubleshoot a device by linking a wiring photo to a schematic and validating the fix with online documentation, or plan a trip by interpreting a transit map and checking schedules under routing constraints. However, existing multimodal benchmarks mainly evaluate single-turn visual reasoning or specific tool skills, and they do not fully capture the realism, visual subtlety, and long-horizon tool use that practical agents require. We introduce AgentVista, a benchmark for generalist multimodal agents that spans 25 sub-domains across 7 categories, pairing realistic and detail-rich visual scenarios with natural hybrid tool use. Tasks require long-horizon tool interactions across modalities, including web search, image search, page navigation, and code-based operations for both image processing and general programming. Comprehensive evaluation of state-of-the-art models exposes significant gaps in their ability to carry out long-horizon multimodal tool use. Even the best model in our evaluation, Gemini-3-Pro with tools, achieves only 27.3% overall accuracy, and hard instances can require more than 25 tool-calling turns. We expect AgentVista to accelerate the development of more capable and reliable multimodal agents for realistic and ultra-challenging problem solving.</p> - https://arxiv.org/abs/2602.23166 - Thu, 26 Feb 2026 16:30:46 +0000 + Mario: Multimodal Graph Reasoning with Large Language Models + https://arxiv.org/abs/2603.05181 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05181.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuanfu Sun, Kang Li, Pengkang Guo, Jiajin Liu, Qiaoyu Tan</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Recent advances in large language models (LLMs) have opened new avenues for multimodal reasoning. Yet, most existing methods still rely on pretrained vision-language models (VLMs) to encode image-text pairs in isolation, ignoring the relational structure that real-world multimodal data naturally form. This motivates reasoning on multimodal graphs (MMGs), where each node has textual and visual attributes and edges provide structural cues. Enabling LLM-based reasoning on such heterogeneous multimodal signals while preserving graph topology introduces two key challenges: resolving weak cross-modal consistency and handling heterogeneous modality preference. To address this, we propose Mario, a unified framework that simultaneously resolves the two above challenges and enables effective LLM-based reasoning over MMGs. Mario consists of two innovative stages. Firstly, a graph-conditioned VLM design that jointly refines textual and visual features through fine-grained cross-modal contrastive learning guided by graph topology. Secondly, a modality-adaptive graph instruction tuning mechanism that organizes aligned multimodal features into graph-aware instruction views and employs a learnable router to surface, for each node and its neighborhood, the most informative modality configuration to the LLM. Extensive experiments across diverse MMG benchmarks demonstrate that Mario consistently outperforms state-of-the-art graph models in both supervised and zero-shot scenarios for node classification and link prediction. The code will be made available at https://github.com/sunyuanfu/Mario.</p> + https://arxiv.org/abs/2603.05181 + Thu, 05 Mar 2026 13:49:41 +0000 - Large Multimodal Models as General In-Context Classifiers - https://arxiv.org/abs/2602.23229 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23229.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Marco Garosi, Matteo Farina, Alessandro Conti, Massimiliano Mancini, Elisa Ricci</p><p><b>Upvotes:</b> 17</p><p><b>Summary:</b> Which multimodal model should we use for classification? Previous studies suggest that the answer lies in CLIP-like contrastive Vision-Language Models (VLMs), due to their remarkable performance in zero-shot classification. In contrast, Large Multimodal Models (LMM) are more suitable for complex tasks. In this work, we argue that this answer overlooks an important capability of LMMs: in-context learning. We benchmark state-of-the-art LMMs on diverse datasets for closed-world classification and find that, although their zero-shot performance is lower than CLIP's, LMMs with a few in-context examples can match or even surpass contrastive VLMs with cache-based adapters, their "in-context" equivalent. We extend this analysis to the open-world setting, where the generative nature of LMMs makes them more suitable for the task. In this challenging scenario, LMMs struggle whenever provided with imperfect context information. To address this issue, we propose CIRCLE, a simple training-free method that assigns pseudo-labels to in-context examples, iteratively refining them with the available context itself. Through extensive experiments, we show that CIRCLE establishes a robust baseline for open-world classification, surpassing VLM counterparts and highlighting the potential of LMMs to serve as unified classifiers, and a flexible alternative to specialized models.</p> - https://arxiv.org/abs/2602.23229 - Thu, 26 Feb 2026 17:08:18 +0000 + Building AI Coding Agents for the Terminal: Scaffolding, Harness, Context Engineering, and Lessons Learned + https://arxiv.org/abs/2603.05344 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05344.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Nghi D. Q. Bui</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> The landscape of AI coding assistance is undergoing a fundamental shift from complex IDE plugins to versatile, terminal-native agents. Operating directly where developers manage source control, execute builds, and deploy environments, CLI-based agents offer unprecedented autonomy for long-horizon development tasks. In this paper, we present OPENDEV, an open-source, command-line coding agent engineered specifically for this new paradigm. Effective autonomous assistance requires strict safety controls and highly efficient context management to prevent context bloat and reasoning degradation. OPENDEV overcomes these challenges through a compound AI system architecture with workload-specialized model routing, a dual-agent architecture separating planning from execution, lazy tool discovery, and adaptive context compaction that progressively reduces older observations. Furthermore, it employs an automated memory system to accumulate project-specific knowledge across sessions and counteracts instruction fade-out through event-driven system reminders. By enforcing explicit reasoning phases and prioritizing context efficiency, OPENDEV provides a secure, extensible foundation for terminal-first AI assistance, offering a blueprint for robust autonomous software engineering.</p> + https://arxiv.org/abs/2603.05344 + Thu, 05 Mar 2026 16:21:08 +0000 - Truncated Step-Level Sampling with Process Rewards for Retrieval-Augmented Reasoning - https://arxiv.org/abs/2602.23440 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23440.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chris Samarinas, Haw-Shiuan Chang, Hamed Zamani</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Training large language models to reason with search engines via reinforcement learning is hindered by a fundamental credit assignment problem: existing methods such as Search-R1 provide only a sparse outcome reward after an entire multi-step trajectory, making it infeasible to attribute success or failure to individual reasoning and retrieval decisions. Process-reward methods like StepSearch alleviate this by introducing step-level supervision, but rely on heuristic rewards such as TF-IDF overlap with gold documents, and still sample k complete trajectories per example, retaining high gradient variance. We propose SLATE, a framework built on two complementary ideas: (1) truncated step-level sampling, which generates k trajectories that share a common prefix and differ only at the next step, and (2) dense LLM-as-judge rewards, which replace heuristic scoring with a capable LLM evaluator that assesses the quality of each reasoning step, search query, and answer, providing richer and more reliable supervision. We theoretically prove that under the same dense reward structure, truncated sampling reduces the variance of advantage estimates by up to a factor of T compared to full-trajectory sampling for T-step trajectories, yielding lower-variance, better-targeted policy gradients. Experiments on seven QA benchmarks confirm that SLATE consistently outperforms both sparse-reward and process-reward baselines, with the largest gains on harder multi-hop tasks and smaller models.</p> - https://arxiv.org/abs/2602.23440 - Thu, 26 Feb 2026 19:05:40 +0000 + Progressive Residual Warmup for Language Model Pretraining + https://arxiv.org/abs/2603.05369 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05369.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tianhao Chen, Xin Xu, Lu Yin, Hao Chen, Yang Wang, Shizhe Diao, Can Yang</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> Transformer architectures serve as the backbone for most modern Large Language Models, therefore their pretraining stability and convergence speed are of central concern. Motivated by the logical dependency of sequentially stacked layers, we propose Progressive Residual Warmup (ProRes) for language model pretraining. ProRes implements an "early layer learns first" philosophy by multiplying each layer's residual with a scalar that gradually warms up from 0 to 1, with deeper layers taking longer warmup steps. In this way, deeper layers wait for early layers to settle into a more stable regime before contributing to learning. We demonstrate the effectiveness of ProRes through pretraining experiments across various model scales, as well as normalization and initialization schemes. Comprehensive analysis shows that ProRes not only stabilizes pretraining but also introduces a unique optimization trajectory, leading to faster convergence, stronger generalization and better downstream performance. Our code is available at https://github.com/dandingsky/ProRes.</p> + https://arxiv.org/abs/2603.05369 + Thu, 05 Mar 2026 16:52:15 +0000 - CiteAudit: You Cited It, But Did You Read It? A Benchmark for Verifying Scientific References in the LLM Era - https://arxiv.org/abs/2602.23452 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23452.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhengqing Yuan, Kaiwen Shi, Zheyuan Zhang, Lichao Sun, Nitesh V. Chawla, Yanfang Ye</p><p><b>Upvotes:</b> 16</p><p><b>Summary:</b> Scientific research relies on accurate citation for attribution and integrity, yet large language models (LLMs) introduce a new risk: fabricated references that appear plausible but correspond to no real publications. Such hallucinated citations have already been observed in submissions and accepted papers at major machine learning venues, exposing vulnerabilities in peer review. Meanwhile, rapidly growing reference lists make manual verification impractical, and existing automated tools remain fragile to noisy and heterogeneous citation formats and lack standardized evaluation. We present the first comprehensive benchmark and detection framework for hallucinated citations in scientific writing. Our multi-agent verification pipeline decomposes citation checking into claim extraction, evidence retrieval, passage matching, reasoning, and calibrated judgment to assess whether a cited source truly supports its claim. We construct a large-scale human-validated dataset across domains and define unified metrics for citation faithfulness and evidence alignment. Experiments with state-of-the-art LLMs reveal substantial citation errors and show that our framework significantly outperforms prior methods in both accuracy and interpretability. This work provides the first scalable infrastructure for auditing citations in the LLM era and practical tools to improve the trustworthiness of scientific references.</p> - https://arxiv.org/abs/2602.23452 - Thu, 26 Feb 2026 19:17:39 +0000 + Planning in 8 Tokens: A Compact Discrete Tokenizer for Latent World Model + https://arxiv.org/abs/2603.05438 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05438.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dongwon Kim, Gawon Seo, Jinsung Lee, Minsu Cho, Suha Kwak</p><p><b>Upvotes:</b> 35</p><p><b>Summary:</b> World models provide a powerful framework for simulating environment dynamics conditioned on actions or instructions, enabling downstream tasks such as action planning or policy learning. Recent approaches leverage world models as learned simulators, but its application to decision-time planning remains computationally prohibitive for real-time control. A key bottleneck lies in latent representations: conventional tokenizers encode each observation into hundreds of tokens, making planning both slow and resource-intensive. To address this, we propose CompACT, a discrete tokenizer that compresses each observation into as few as 8 tokens, drastically reducing computational cost while preserving essential information for planning. An action-conditioned world model that occupies CompACT tokenizer achieves competitive planning performance with orders-of-magnitude faster planning, offering a practical step toward real-world deployment of world models.</p> + https://arxiv.org/abs/2603.05438 + Thu, 05 Mar 2026 18:00:02 +0000 - Humans and LLMs Diverge on Probabilistic Inferences - https://arxiv.org/abs/2602.23546 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23546.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Gaurav Kamath, Sreenath Madathil, Sebastian Schuster, Marie-Catherine de Marneffe, Siva Reddy</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Human reasoning often involves working over limited information to arrive at probabilistic conclusions. In its simplest form, this involves making an inference that is not strictly entailed by a premise, but rather only likely given the premise. While reasoning LLMs have demonstrated strong performance on logical and mathematical tasks, their behavior on such open-ended, non-deterministic inferences remains largely unexplored. We introduce ProbCOPA, a dataset of 210 handcrafted probabilistic inferences in English, each annotated for inference likelihood by 25--30 human participants. We find that human responses are graded and varied, revealing probabilistic judgments of the inferences in our dataset. Comparing these judgments with responses from eight state-of-the-art reasoning LLMs, we show that models consistently fail to produce human-like distributions. Finally, analyzing LLM reasoning chains, we find evidence of a common reasoning pattern used to evaluate such inferences. Our findings reveal persistent differences between humans and LLMs, and underscore the need to evaluate reasoning beyond deterministic settings.</p> - https://arxiv.org/abs/2602.23546 - Thu, 26 Feb 2026 23:00:41 +0000 + Scaling Agentic Capabilities, Not Context: Efficient Reinforcement Finetuning for Large Toolspaces + https://arxiv.org/abs/2603.06713 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06713.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Karan Gupta, Pranav Vajreshwari, Yash Pandya, Raghav Magazine, Akshay Nambi, Ahmed Awadallah</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> Agentic systems operating over large tool ecosystems must plan and execute long-horizon workflows under weak or non-verifiable supervision. While frontier models mitigate these challenges through scale and large context budgets, small language models (SLMs) remain brittle: eager tool loading saturates context, execution errors compound over time, and sparse rewards limit learning. We introduce ATLAS, a reinforcement finetuning framework that enables SLMs to operate effectively in large-scale toolspace environments by learning how to acquire context and how to execute actions. Our approach makes two key contributions. First, we treat context control and execution structure as learnable decisions, combining iterative tool loading with programmatic tool orchestration to bound context growth and stabilize long-horizon trajectories. Second, we propose rubric-based reinforcement finetuning, which decomposes task success into structured, task-aligned criteria and enables scalable training using small judge models. Across MCP benchmarks, these design choices yield large and consistent gains over generic RL baselines, allowing a 4B SLM to approach frontier-agent performance under far tighter parameter and context budgets.</p> + https://arxiv.org/abs/2603.06713 + Thu, 05 Mar 2026 20:29:12 +0000 - SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale - https://arxiv.org/abs/2602.23866 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23866.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ibragim Badertdinov, Maksim Nekrashevich, Anton Shevtsov, Alexander Golubev</p><p><b>Upvotes:</b> 76</p><p><b>Summary:</b> Software engineering agents (SWE) are improving rapidly, with recent gains largely driven by reinforcement learning (RL). However, RL training is constrained by the scarcity of large-scale task collections with reproducible execution environments and reliable test suites. Although a growing number of benchmarks have emerged, datasets suitable for training remain limited in scale and diversity or often target a limited set of high-resource language ecosystems. We introduce SWE-rebench V2, a language-agnostic automated pipeline for harvesting executable real-world SWE tasks and constructing RL training environments at scale. The pipeline synthesizes repository-specific installation and test procedures via an interactive setup agent, and filters unsound instances using an ensemble of LLM judges, validated against human-verified SWE-bench annotations. Using this pipeline, we construct a dataset of 32,000+ tasks spanning 20 languages and 3,600+ repositories, with pre-built images for reproducible execution. To further scale training data, we additionally release 120,000+ tasks with installation instructions, fail-to-pass tests and rich metadata, where the problem statement is generated based on the original pull request description. We validate the collected instances through a diagnostic study that covers a subset of tasks in five programming languages across seven popular models, and provide instance-level metadata that flags common confounders such as overly restrictive tests and underspecified descriptions. We release the datasets, the collection and execution code, and associated artifacts to enable large-scale training of SWE agents across diverse languages and repositories.</p> - https://arxiv.org/abs/2602.23866 - Fri, 27 Feb 2026 10:06:10 +0000 + Reasoning Models Struggle to Control their Chains of Thought + https://arxiv.org/abs/2603.05706 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05706.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chen Yueh-Han, Robert McCarthy, Bruce W. Lee, He He, Ian Kivlichan, Bowen Baker, Micah Carroll, Tomek Korbak</p><p><b>Upvotes:</b> 27</p><p><b>Summary:</b> Chain-of-thought (CoT) monitoring is a promising tool for detecting misbehaviors and understanding the motivations of modern reasoning models. However, if models can control what they verbalize in their CoT, it could undermine CoT monitorability. To measure this undesirable capability -- CoT controllability -- we introduce the CoT-Control evaluation suite, which includes tasks that require models to solve problems while adhering to CoT instructions, e.g., reasoning about a genetics question without using the word 'chromosome'. We show that reasoning models possess significantly lower CoT controllability than output controllability; for instance, Claude Sonnet 4.5 can control its CoT only 2.7% of the time but 61.9% when controlling its final output. We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty. CoT controllability failures extend even to situations in which models are given incentives (as opposed to direct requests) to evade CoT monitors, although models exhibit slightly higher controllability when they are told they are being monitored. Similarly, eliciting controllability by adversarially optimizing prompts does not meaningfully increase controllability. Our results leave us cautiously optimistic that CoT controllability is currently unlikely to be a failure mode of CoT monitorability. However, the mechanism behind low controllability is not well understood. Given its importance for maintaining CoT monitorability, we recommend that frontier labs track CoT controllability in future models.</p> + https://arxiv.org/abs/2603.05706 + Thu, 05 Mar 2026 22:03:48 +0000 - LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding - https://arxiv.org/abs/2602.23881 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23881.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Alexander Samarin, Sergei Krutikov, Anton Shevtsov, Sergei Skvortsov, Filipp Fisin, Alexander Golubev</p><p><b>Upvotes:</b> 18</p><p><b>Summary:</b> Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to propose candidate tokens that are then verified in parallel by the target model. The speedup is significantly determined by the acceptance rate, yet standard training minimizes Kullback-Leibler (KL) divergence as a proxy objective. While KL divergence and acceptance rate share the same global optimum, small draft models, having limited capacity, typically converge to suboptimal solutions where minimizing KL does not guarantee maximizing acceptance rate. To address this issue, we propose LK losses, special training objectives that directly target acceptance rate. Comprehensive experiments across four draft architectures and six target models, ranging from 8B to 685B parameters, demonstrate consistent improvements in acceptance metrics across all configurations compared to the standard KL-based training. We evaluate our approach on general, coding and math domains and report gains of up to 8-10% in average acceptance length. LK losses are easy to implement, introduce no computational overhead and can be directly integrated into any existing speculator training framework, making them a compelling alternative to the existing draft training objectives.</p> - https://arxiv.org/abs/2602.23881 - Fri, 27 Feb 2026 10:20:11 +0000 + Training-free Latent Inter-Frame Pruning with Attention Recovery + https://arxiv.org/abs/2603.05811 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05811.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dennis Menn, Yuedong Yang, Bokun Wang, Xiwen Wei, Mustafa Munir, Feng Liang, Radu Marculescu, Chenfeng Xu, Diana Marculescu</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Current video generation models suffer from high computational latency, making real-time applications prohibitively costly. In this paper, we address this limitation by exploiting the temporal redundancy inherent in video latent patches. To this end, we propose the Latent Inter-frame Pruning with Attention Recovery (LIPAR) framework, which detects and skips recomputing duplicated latent patches. Additionally, we introduce a novel Attention Recovery mechanism that approximates the attention values of pruned tokens, thereby removing visual artifacts arising from naively applying the pruning method. Empirically, our method increases video editing throughput by 1.45times, on average achieving 12.2 FPS on an NVIDIA A6000 compared to the baseline 8.4 FPS. The proposed method does not compromise generation quality and can be seamlessly integrated with the model without additional training. Our approach effectively bridges the gap between traditional compression algorithms and modern generative pipelines.</p> + https://arxiv.org/abs/2603.05811 + Fri, 06 Mar 2026 01:49:47 +0000 - Ref-Adv: Exploring MLLM Visual Reasoning in Referring Expression Tasks - https://arxiv.org/abs/2602.23898 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23898.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qihua Dong, Kuo Yang, Lin Ju, Handong Zhao, Yitian Zhang, Yizhou Wang, Huimin Zeng, Jianglin Lu, Yun Fu</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Referring Expression Comprehension (REC) links language to region level visual perception. Standard benchmarks (RefCOCO, RefCOCO+, RefCOCOg) have progressed rapidly with multimodal LLMs but remain weak tests of visual reasoning and grounding: (i) many expressions are very short, leaving little reasoning demand; (ii) images often contain few distractors, making the target easy to find; and (iii) redundant descriptors enable shortcut solutions that bypass genuine text understanding and visual reasoning. We introduce Ref-Adv, a modern REC benchmark that suppresses shortcuts by pairing linguistically nontrivial expressions with only the information necessary to uniquely identify the target. The dataset contains referring expressions on real images, curated with hard distractors and annotated with reasoning facets including negation. We conduct comprehensive ablations (word order perturbations and descriptor deletion sufficiency) to show that solving Ref-Adv requires reasoning beyond simple cues, and we evaluate a broad suite of contemporary multimodal LLMs on Ref-Adv. Despite strong results on RefCOCO, RefCOCO+, and RefCOCOg, models drop markedly on Ref-Adv, revealing reliance on shortcuts and gaps in visual reasoning and grounding. We provide an in depth failure analysis and aim for Ref-Adv to guide future work on visual reasoning and grounding in MLLMs.</p> - https://arxiv.org/abs/2602.23898 - Fri, 27 Feb 2026 10:47:26 +0000 + ReflexiCoder: Teaching Large Language Models to Self-Reflect on Generated Code and Self-Correct It via Reinforcement Learning + https://arxiv.org/abs/2603.05863 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05863.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Juyong Jiang, Jiasi Shen, Sunghun Kim, Kang Min Yoo, Jeonghoon Kim, Sungju Kim</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> While Large Language Models (LLMs) have revolutionized code generation, standard "System 1" approaches, generating solutions in a single forward pass, often hit a performance ceiling when faced with complex algorithmic tasks. Existing iterative refinement strategies attempt to bridge this gap at inference time, yet they predominantly rely on external oracles, execution feedback, or computationally expensive prompt-response cycles. In this work, we propose ReflexiCoder, a novel reinforcement learning (RL) framework that internalizes the structured reasoning trajectory, encompassing initial generation, bug and optimization aware reflection, and self-correction, directly into the model's weights. Unlike prior methods, ReflexiCoder shifts the paradigm from external-dependent refinement to an intrinsic, fully autonomous self-reflection and self-correction capabilities at inference time. We utilize an RL-zero training paradigm with granular reward functions to optimize the entire reflection-correction trajectory, teaching the model how to debug without reliance on ground-truth feedback or execution engines at inference time. Extensive experiments across seven benchmarks demonstrate that our ReflexiCoder-8B establishes a new state-of-the-art (SOTA) among leading open-source models in the 1.5B-14B range, achieving 94.51% (87.20%) on HumanEval (Plus), 81.80% (78.57%) on MBPP (Plus), 35.00% on BigCodeBench, 52.21% on LiveCodeBench, and 37.34% on CodeForces in a single-attempt setting, rivaling or surpassing proprietary models like GPT-5.1. Notably, our framework is significantly more token-efficient than base models, reducing inference-time compute overhead by approximately 40% through disciplined, high-speed reasoning and reflection patterns. Source code is available at https://github.com/juyongjiang/ReflexiCoder.</p> + https://arxiv.org/abs/2603.05863 + Fri, 06 Mar 2026 03:38:17 +0000 - Half-Truths Break Similarity-Based Retrieval - https://arxiv.org/abs/2602.23906 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23906.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Bora Kargi, Arnas Uselis, Seong Joon Oh</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> When a text description is extended with an additional detail, image-text similarity should drop if that detail is wrong. We show that CLIP-style dual encoders often violate this intuition: appending a plausible but incorrect object or relation to an otherwise correct description can increase the similarity score. We call such cases half-truths. On COCO, CLIP prefers the correct shorter description only 40.6% of the time, and performance drops to 32.9% when the added detail is a relation. We trace this vulnerability to weak supervision on caption parts: contrastive training aligns full sentences but does not explicitly enforce that individual entities and relations are grounded. We propose CS-CLIP (Component-Supervised CLIP), which decomposes captions into entity and relation units, constructs a minimally edited foil for each unit, and fine-tunes the model to score the correct unit above its foil while preserving standard dual-encoder inference. CS-CLIP raises half-truth accuracy to 69.3% and improves average performance on established compositional benchmarks by 5.7 points, suggesting that reducing half-truth errors aligns with broader gains in compositional understanding. Code is publicly available at: https://github.com/kargibora/CS-CLIP</p> - https://arxiv.org/abs/2602.23906 - Fri, 27 Feb 2026 10:56:31 +0000 + Lost in Stories: Consistency Bugs in Long Story Generation by LLMs + https://arxiv.org/abs/2603.05890 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05890.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junjie Li, Xinrui Guo, Yuhao Wu, Roy Ka-Wei Lee, Hongzhi Li, Yutao Xie</p><p><b>Upvotes:</b> 82</p><p><b>Summary:</b> What happens when a storyteller forgets its own story? Large Language Models (LLMs) can now generate narratives spanning tens of thousands of words, but they often fail to maintain consistency throughout. When generating long-form narratives, these models can contradict their own established facts, character traits, and world rules. Existing story generation benchmarks focus mainly on plot quality and fluency, leaving consistency errors largely unexplored. To address this gap, we present ConStory-Bench, a benchmark designed to evaluate narrative consistency in long-form story generation. It contains 2,000 prompts across four task scenarios and defines a taxonomy of five error categories with 19 fine-grained subtypes. We also develop ConStory-Checker, an automated pipeline that detects contradictions and grounds each judgment in explicit textual evidence. Evaluating a range of LLMs through five research questions, we find that consistency errors show clear tendencies: they are most common in factual and temporal dimensions, tend to appear around the middle of narratives, occur in text segments with higher token-level entropy, and certain error types tend to co-occur. These findings can inform future efforts to improve consistency in long-form narrative generation. Our project page is available at https://picrew.github.io/constory-bench.github.io/.</p> + https://arxiv.org/abs/2603.05890 + Fri, 06 Mar 2026 04:16:46 +0000 - Accelerating Masked Image Generation by Learning Latent Controlled Dynamics - https://arxiv.org/abs/2602.23996 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.23996.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kaiwen Zhu, Quansheng Zeng, Yuandong Pu, Shuo Cao, Xiaohui Li, Yi Xin, Qi Qin, Jiayang Li, Yu Qiao, Jinjin Gu, Yihao Liu</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Masked Image Generation Models (MIGMs) have achieved great success, yet their efficiency is hampered by the multiple steps of bi-directional attention. In fact, there exists notable redundancy in their computation: when sampling discrete tokens, the rich semantics contained in the continuous features are lost. Some existing works attempt to cache the features to approximate future features. However, they exhibit considerable approximation error under aggressive acceleration rates. We attribute this to their limited expressivity and the failure to account for sampling information. To fill this gap, we propose to learn a lightweight model that incorporates both previous features and sampled tokens, and regresses the average velocity field of feature evolution. The model has moderate complexity that suffices to capture the subtle dynamics while keeping lightweight compared to the original base model. We apply our method, MIGM-Shortcut, to two representative MIGM architectures and tasks. In particular, on the state-of-the-art Lumina-DiMOO, it achieves over 4x acceleration of text-to-image generation while maintaining quality, significantly pushing the Pareto frontier of masked image generation. The code and model weights are available at https://github.com/Kaiwen-Zhu/MIGM-Shortcut.</p> - https://arxiv.org/abs/2602.23996 - Fri, 27 Feb 2026 13:16:58 +0000 + EffectMaker: Unifying Reasoning and Generation for Customized Visual Effect Creation + https://arxiv.org/abs/2603.06014 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06014.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shiyuan Yang, Ruihuang Li, Jiale Tao, Shuai Shao, Qinglin Lu, Jing Liao</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Visual effects (VFX) are essential for enhancing the expressiveness and creativity of video content, yet producing high-quality effects typically requires expert knowledge and costly production pipelines. Existing AIGC systems face significant challenges in VFX generation due to the scarcity of effect-specific data and the inherent difficulty of modeling supernatural or stylized effects. Moreover, these approaches often require per-effect fine-tuning, which severely limits their scalability and generalization to novel VFX. In this work, we present EffectMaker, a unified reasoning-generation framework that enables reference-based VFX customization. EffectMaker employs a multimodal large language model to interpret high-level effect semantics and reason about how they should adapt to a target subject, while a diffusion transformer leverages in-context learning to capture fine-grained visual cues from reference videos. These two components form a semantic-visual dual-path guidance mechanism that enables accurate, controllable, and effect-consistent synthesis without per-effect fine-tuning. Furthermore, we construct EffectData, the largest high-quality synthetic dataset containing 130k videos across 3k VFX categories, to improve generalization and scalability. Experiments show that EffectMaker achieves superior visual quality and effect consistency over state-of-the-art baselines, offering a scalable and flexible paradigm for customized VFX generation. Project page: https://effectmaker.github.io</p> + https://arxiv.org/abs/2603.06014 + Fri, 06 Mar 2026 08:09:14 +0000 - InfoNCE Induces Gaussian Distribution - https://arxiv.org/abs/2602.24012 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24012.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Roy Betser, Eyal Gofer, Meir Yossef Levi, Guy Gilboa</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Contrastive learning has become a cornerstone of modern representation learning, allowing training with massive unlabeled data for both task-specific and general (foundation) models. A prototypical loss in contrastive training is InfoNCE and its variants. In this work, we show that the InfoNCE objective induces Gaussian structure in representations that emerge from contrastive training. We establish this result in two complementary regimes. First, we show that under certain alignment and concentration assumptions, projections of the high-dimensional representation asymptotically approach a multivariate Gaussian distribution. Next, under less strict assumptions, we show that adding a small asymptotically vanishing regularization term that promotes low feature norm and high feature entropy leads to similar asymptotic results. We support our analysis with experiments on synthetic and CIFAR-10 datasets across multiple encoder architectures and sizes, demonstrating consistent Gaussian behavior. This perspective provides a principled explanation for commonly observed Gaussianity in contrastive representations. The resulting Gaussian model enables principled analytical treatment of learned representations and is expected to support a wide range of applications in contrastive learning.</p> - https://arxiv.org/abs/2602.24012 - Fri, 27 Feb 2026 13:35:58 +0000 + FlashPrefill: Instantaneous Pattern Discovery and Thresholding for Ultra-Fast Long-Context Prefilling + https://arxiv.org/abs/2603.06199 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06199.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qihang Fan, Huaibo Huang, Zhiying Wu, Juqiu Wang, Bingning Wang, Ran He</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Long-context modeling is a pivotal capability for Large Language Models, yet the quadratic complexity of attention remains a critical bottleneck, particularly during the compute-intensive prefilling phase. While various sparse attention mechanisms have been explored, they typically suffer from either significant search latency or insufficient sparsity. In this paper, we propose FlashPrefill, a framework enabling ultra-fast prefilling via instantaneous pattern discovery and thresholding. FlashPrefill leverages a fast block-searching technique to simultaneously locate dynamic vertical, slash, and block-sparse attention patterns. Crucially, it introduces a dynamic thresholding mechanism that bypasses the prohibitive overhead of sorting or accumulating attention scores while effectively eliminating the long-tail distribution to enhance sparsity. Extensive evaluations demonstrate that FlashPrefill achieves a substantial leap in efficiency, delivering an unprecedented 27.78x speedup on 256K sequences. Notably, unlike existing methods that incur efficiency degradation on shorter contexts, FlashPrefill maintains a 1.71x speedup even at a 4K context length, demonstrating its robustness and practical utility across varying sequence scales.</p> + https://arxiv.org/abs/2603.06199 + Fri, 06 Mar 2026 12:12:46 +0000 - SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching - https://arxiv.org/abs/2602.24208 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24208.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yasaman Haghighi, Alexandre Alahi</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Diffusion models achieve state-of-the-art video generation quality, but their inference remains expensive due to the large number of sequential denoising steps. This has motivated a growing line of research on accelerating diffusion inference. Among training-free acceleration methods, caching reduces computation by reusing previously computed model outputs across timesteps. Existing caching methods rely on heuristic criteria to choose cache/reuse timesteps and require extensive tuning. We address this limitation with a principled sensitivity-aware caching framework. Specifically, we formalize the caching error through an analysis of the model output sensitivity to perturbations in the denoising inputs, i.e., the noisy latent and the timestep, and show that this sensitivity is a key predictor of caching error. Based on this analysis, we propose Sensitivity-Aware Caching (SenCache), a dynamic caching policy that adaptively selects caching timesteps on a per-sample basis. Our framework provides a theoretical basis for adaptive caching, explains why prior empirical heuristics can be partially effective, and extends them to a dynamic, sample-specific approach. Experiments on Wan 2.1, CogVideoX, and LTX-Video show that SenCache achieves better visual quality than existing caching methods under similar computational budgets.</p> - https://arxiv.org/abs/2602.24208 - Fri, 27 Feb 2026 17:36:09 +0000 + Dynamic Chunking Diffusion Transformer + https://arxiv.org/abs/2603.06351 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06351.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Akash Haridas, Utkarsh Saxena, Parsa Ashrafi Fashi, Mehdi Rezagholizadeh, Vikram Appia, Emad Barsoum</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Diffusion Transformers process images as fixed-length sequences of tokens produced by a static patchify operation. While effective, this design spends uniform compute on low- and high-information regions alike, ignoring that images contain regions of varying detail and that the denoising process progresses from coarse structure at early timesteps to fine detail at late timesteps. We introduce the Dynamic Chunking Diffusion Transformer (DC-DiT), which augments the DiT backbone with a learned encoder-router-decoder scaffold that adaptively compresses the 2D input into a shorter token sequence in a data-dependent manner using a chunking mechanism learned end-to-end with diffusion training. The mechanism learns to compress uniform background regions into fewer tokens and detail-rich regions into more tokens, with meaningful visual segmentations emerging without explicit supervision. Furthermore, it also learns to adapt its compression across diffusion timesteps, using fewer tokens at noisy stages and more tokens as fine details emerge. On class-conditional ImageNet 256{times}256, DC-DiT consistently improves FID and Inception Score over both parameter-matched and FLOP-matched DiT baselines across 4{times} and 16{times} compression, showing this is a promising technique with potential further applications to pixel-space, video and 3D generation. Beyond accuracy, DC-DiT is practical: it can be upcycled from pretrained DiT checkpoints with minimal post-training compute (up to 8{times} fewer training steps) and composes with other dynamic computation methods to further reduce generation FLOPs.</p> + https://arxiv.org/abs/2603.06351 + Fri, 06 Mar 2026 14:59:11 +0000 - Enhancing Spatial Understanding in Image Generation via Reward Modeling - https://arxiv.org/abs/2602.24233 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24233.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhenyu Tang, Chaoran Feng, Yufan Deng, Jie Wu, Xiaojie Li, Rui Wang, Yunpeng Chen, Daquan Zhou</p><p><b>Upvotes:</b> 49</p><p><b>Summary:</b> Recent progress in text-to-image generation has greatly advanced visual fidelity and creativity, but it has also imposed higher demands on prompt complexity-particularly in encoding intricate spatial relationships. In such cases, achieving satisfactory results often requires multiple sampling attempts. To address this challenge, we introduce a novel method that strengthens the spatial understanding of current image generation models. We first construct the SpatialReward-Dataset with over 80k preference pairs. Building on this dataset, we build SpatialScore, a reward model designed to evaluate the accuracy of spatial relationships in text-to-image generation, achieving performance that even surpasses leading proprietary models on spatial evaluation. We further demonstrate that this reward model effectively enables online reinforcement learning for the complex spatial generation. Extensive experiments across multiple benchmarks show that our specialized reward model yields significant and consistent gains in spatial understanding for image generation.</p> - https://arxiv.org/abs/2602.24233 - Fri, 27 Feb 2026 17:59:57 +0000 + Physical Simulator In-the-Loop Video Generation + https://arxiv.org/abs/2603.06408 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06408.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lin Geng Foo, Mark He Huang, Alexandros Lattas, Stylianos Moschoglou, Thabo Beeler, Christian Theobalt</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Recent advances in diffusion-based video generation have achieved remarkable visual realism but still struggle to obey basic physical laws such as gravity, inertia, and collision. Generated objects often move inconsistently across frames, exhibit implausible dynamics, or violate physical constraints, limiting the realism and reliability of AI-generated videos. We address this gap by introducing Physical Simulator In-the-loop Video Generation (PSIVG), a novel framework that integrates a physical simulator into the video diffusion process. Starting from a template video generated by a pre-trained diffusion model, PSIVG reconstructs the 4D scene and foreground object meshes, initializes them within a physical simulator, and generates physically consistent trajectories. These simulated trajectories are then used to guide the video generator toward spatio-temporally physically coherent motion. To further improve texture consistency during object movement, we propose a Test-Time Texture Consistency Optimization (TTCO) technique that adapts text and feature embeddings based on pixel correspondences from the simulator. Comprehensive experiments demonstrate that PSIVG produces videos that better adhere to real-world physics while preserving visual quality and diversity. Project Page: https://vcai.mpi-inf.mpg.de/projects/PSIVG/</p> + https://arxiv.org/abs/2603.06408 + Fri, 06 Mar 2026 15:48:25 +0000 - Compositional Generalization Requires Linear, Orthogonal Representations in Vision Embedding Models - https://arxiv.org/abs/2602.24264 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24264.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Arnas Uselis, Andrea Dittadi, Seong Joon Oh</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Compositional generalization, the ability to recognize familiar parts in novel contexts, is a defining property of intelligent systems. Although modern models are trained on massive datasets, they still cover only a tiny fraction of the combinatorial space of possible inputs, raising the question of what structure representations must have to support generalization to unseen combinations. We formalize three desiderata for compositional generalization under standard training (divisibility, transferability, stability) and show they impose necessary geometric constraints: representations must decompose linearly into per-concept components, and these components must be orthogonal across concepts. This provides theoretical grounding for the Linear Representation Hypothesis: the linear structure widely observed in neural representations is a necessary consequence of compositional generalization. We further derive dimension bounds linking the number of composable concepts to the embedding geometry. Empirically, we evaluate these predictions across modern vision models (CLIP, SigLIP, DINO) and find that representations exhibit partial linear factorization with low-rank, near-orthogonal per-concept factors, and that the degree of this structure correlates with compositional generalization on unseen combinations. As models continue to scale, these conditions predict the representational geometry they may converge to. Code is available at https://github.com/oshapio/necessary-compositionality.</p> - https://arxiv.org/abs/2602.24264 - Fri, 27 Feb 2026 18:32:31 +0000 + CaTok: Taming Mean Flows for One-Dimensional Causal Image Tokenization + https://arxiv.org/abs/2603.06449 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06449.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yitong Chen, Zuxuan Wu, Xipeng Qiu, Yu-Gang Jiang</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Autoregressive (AR) language models rely on causal tokenization, but extending this paradigm to vision remains non-trivial. Current visual tokenizers either flatten 2D patches into non-causal sequences or enforce heuristic orderings that misalign with the "next-token prediction" pattern. Recent diffusion autoencoders similarly fall short: conditioning the decoder on all tokens lacks causality, while applying nested dropout mechanism introduces imbalance. To address these challenges, we present CaTok, a 1D causal image tokenizer with a MeanFlow decoder. By selecting tokens over time intervals and binding them to the MeanFlow objective, as illustrated in Fig. 1, CaTok learns causal 1D representations that support both fast one-step generation and high-fidelity multi-step sampling, while naturally capturing diverse visual concepts across token intervals. To further stabilize and accelerate training, we propose a straightforward regularization REPA-A, which aligns encoder features with Vision Foundation Models (VFMs). Experiments demonstrate that CaTok achieves state-of-the-art results on ImageNet reconstruction, reaching 0.75 FID, 22.53 PSNR and 0.674 SSIM with fewer training epochs, and the AR model attains performance comparable to leading approaches.</p> + https://arxiv.org/abs/2603.06449 + Fri, 06 Mar 2026 16:39:17 +0000 - Memory Caching: RNNs with Growing Memory - https://arxiv.org/abs/2602.24281 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24281.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ali Behrouz, Zeman Li, Yuan Deng, Peilin Zhong, Meisam Razaviyayn, Vahab Mirrokni</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Transformers have been established as the de-facto backbones for most recent advances in sequence modeling, mainly due to their growing memory capacity that scales with the context length. While plausible for retrieval tasks, it causes quadratic complexity and so has motivated recent studies to explore viable subquadratic recurrent alternatives. Despite showing promising preliminary results in diverse domains, such recurrent architectures underperform Transformers in recall-intensive tasks, often attributed to their fixed-size memory. In this paper, we introduce Memory Caching (MC), a simple yet effective technique that enhances recurrent models by caching checkpoints of their memory states (a.k.a. hidden states). Memory Caching allows the effective memory capacity of RNNs to grow with sequence length, offering a flexible trade-off that interpolates between the fixed memory (i.e., O(L) complexity) of RNNs and the growing memory (i.e., O(L^2) complexity) of Transformers. We propose four variants of MC, including gated aggregation and sparse selective mechanisms, and discuss their implications on both linear and deep memory modules. Our experimental results on language modeling, and long-context understanding tasks show that MC enhances the performance of recurrent models, supporting its effectiveness. The results of in-context recall tasks indicate that while Transformers achieve the best accuracy, our MC variants show competitive performance, close the gap with Transformers, and performs better than state-of-the-art recurrent models.</p> - https://arxiv.org/abs/2602.24281 - Fri, 27 Feb 2026 18:53:41 +0000 + Penguin-VL: Exploring the Efficiency Limits of VLM with LLM-based Vision Encoders + https://arxiv.org/abs/2603.06569 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06569.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Boqiang Zhang, Lei Ke, Ruihan Yang, Qi Gao, Tianyuan Qu, Rossell Chen, Dong Yu, Leoweiliang</p><p><b>Upvotes:</b> 104</p><p><b>Summary:</b> Vision Language Model (VLM) development has largely relied on scaling model size, which hinders deployment on compute-constrained mobile and edge devices such as smartphones and robots. In this work, we explore the performance limits of compact (e.g., 2B and 8B) VLMs. We challenge the prevailing practice that state-of-the-art VLMs must rely on vision encoders initialized via massive contrastive pretraining (e.g., CLIP/SigLIP). We identify an objective mismatch: contrastive learning, optimized for discrimination, enforces coarse and category-level invariances that suppress fine-grained visual cues needed for dense captioning and complex VLM reasoning. To address this issue, we present Penguin-VL, whose vision encoder is initialized from a text-only LLM. Our experiments reveal that Penguin-Encoder serves as a superior alternative to traditional contrastive pretraining, unlocking a higher degree of visual fidelity and data efficiency for multimodal understanding. Across various image and video benchmarks, Penguin-VL achieves performance comparable to leading VLMs (e.g., Qwen3-VL) in mathematical reasoning and surpasses them in tasks such as document understanding, visual knowledge, and multi-perspective video understanding. Notably, these gains are achieved with a lightweight architecture, demonstrating that improved visual representation rather than model scaling is the primary driver of performance. Our ablations show that Penguin-Encoder consistently outperforms contrastive-pretrained encoders, preserving fine-grained spatial and temporal cues that are critical for dense perception and complex reasoning. This makes it a strong drop-in alternative for compute-efficient VLMs and enables high performance in resource-constrained settings. Code: https://github.com/tencent-ailab/Penguin-VL</p> + https://arxiv.org/abs/2603.06569 + Fri, 06 Mar 2026 18:58:04 +0000 - CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation - https://arxiv.org/abs/2602.24286 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24286.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Weinan Dai, Hanlin Wu, Qiying Yu, Huan-ang Gao, Jiahao Li, Chengquan Jiang, Weiqiang Lou, Yufan Song, Hongli Yu, Jiaze Chen, Wei-Ying Ma, Ya-Qin Zhang, Jingjing Liu, Mingxuan Wang, Xin Liu, Hao Zhou</p><p><b>Upvotes:</b> 79</p><p><b>Summary:</b> GPU kernel optimization is fundamental to modern deep learning but remains a highly specialized task requiring deep hardware expertise. Despite strong performance in general programming, large language models (LLMs) remain uncompetitive with compiler-based systems such as torch.compile for CUDA kernel generation. Existing CUDA code generation approaches either rely on training-free refinement or fine-tune models within fixed multi-turn execution-feedback loops, but both paradigms fail to fundamentally improve the model's intrinsic CUDA optimization ability, resulting in limited performance gains. We present CUDA Agent, a large-scale agentic reinforcement learning system that develops CUDA kernel expertise through three components: a scalable data synthesis pipeline, a skill-augmented CUDA development environment with automated verification and profiling to provide reliable reward signals, and reinforcement learning algorithmic techniques enabling stable training. CUDA Agent achieves state-of-the-art results on KernelBench, delivering 100\%, 100\%, and 92\% faster rate over torch.compile on KernelBench Level-1, Level-2, and Level-3 splits, outperforming the strongest proprietary models such as Claude Opus 4.5 and Gemini 3 Pro by about 40\% on the hardest Level-3 setting.</p> - https://arxiv.org/abs/2602.24286 - Fri, 27 Feb 2026 18:58:05 +0000 + Omni-Diffusion: Unified Multimodal Understanding and Generation with Masked Discrete Diffusion + https://arxiv.org/abs/2603.06577 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06577.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lijiang Li, Zuwei Long, Yunhang Shen, Heting Gao, Haoyu Cao, Xing Sun, Caifeng Shan, Ran He, Chaoyou Fu</p><p><b>Upvotes:</b> 43</p><p><b>Summary:</b> While recent multimodal large language models (MLLMs) have made impressive strides, they predominantly employ a conventional autoregressive architecture as their backbone, leaving significant room to explore effective and efficient alternatives in architectural design. Concurrently, recent studies have successfully applied discrete diffusion models to various domains, such as visual understanding and image generation, revealing their considerable potential as a promising backbone for multimodal systems. Drawing inspiration from these pioneering research, we introduce Omni-Diffusion, the first any-to-any multimodal language model built entirely on mask-based discrete diffusion models, which unifies understanding and generation across text, speech, and images. Omni-Diffusion employs a unified mask-based discrete diffusion model to directly capture the joint distribution over discrete multimodal tokens. This approach supports not only bimodal tasks but also more complex scenarios involving multiple modalities. On a diverse set of benchmarks, our method outperforms or performs on par with existing multimodal systems that process two or more modalities, highlighting the significant promise of diffusion models in powering the next generation of multimodal foundation models. Project webpage: https://omni-diffusion.github.io.</p> + https://arxiv.org/abs/2603.06577 + Fri, 06 Mar 2026 18:59:57 +0000 - Mode Seeking meets Mean Seeking for Fast Long Video Generation - https://arxiv.org/abs/2602.24289 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.24289.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shengqu Cai, Weili Nie, Chao Liu, Julius Berner, Lvmin Zhang, Nanye Ma, Hansheng Chen, Maneesh Agrawala, Leonidas Guibas, Gordon Wetzstein, Arash Vahdat</p><p><b>Upvotes:</b> 38</p><p><b>Summary:</b> Scaling video generation from seconds to minutes faces a critical bottleneck: while short-video data is abundant and high-fidelity, coherent long-form data is scarce and limited to narrow domains. To address this, we propose a training paradigm where Mode Seeking meets Mean Seeking, decoupling local fidelity from long-term coherence based on a unified representation via a Decoupled Diffusion Transformer. Our approach utilizes a global Flow Matching head trained via supervised learning on long videos to capture narrative structure, while simultaneously employing a local Distribution Matching head that aligns sliding windows to a frozen short-video teacher via a mode-seeking reverse-KL divergence. This strategy enables the synthesis of minute-scale videos that learns long-range coherence and motions from limited long videos via supervised flow matching, while inheriting local realism by aligning every sliding-window segment of the student to a frozen short-video teacher, resulting in a few-step fast long video generator. Evaluations show that our method effectively closes the fidelity-horizon gap by jointly improving local sharpness, motion and long-range consistency. Project website: https://primecai.github.io/mmm/.</p> - https://arxiv.org/abs/2602.24289 - Fri, 27 Feb 2026 18:59:02 +0000 + Are Audio-Language Models Listening? Audio-Specialist Heads for Adaptive Audio Steering + https://arxiv.org/abs/2603.06854 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.06854.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Neta Glazer, Lenny Aharon, Ethan Fetaya</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Multimodal large language models can exhibit text dominance, over-relying on linguistic priors instead of grounding predictions in non-text inputs. One example is large audio-language models (LALMs) where decisive audio evidence can be under-utilized even when it contains important information. To address this issue we use mechanistic interpretability to identify a small set of audio-specialist attention heads whose audio attention yields a ``listening'' signal. We show that this signal increases when audio evidence affects the model's output, providing an indicator of audio engagement under standard prompting. Leveraging this localization, we construct an audio--silence steering direction and apply an inference-time activation intervention to the final representation, amplifying the model's audio effect. To demonstrate the utility of this intervention, we show on MMAU that this improves accuracy by up to +8.0 percentage points on two Qwen-based LALMs, without any parameter updates.</p> + https://arxiv.org/abs/2603.06854 + Fri, 06 Mar 2026 20:18:30 +0000 - DreamWorld: Unified World Modeling in Video Generation - https://arxiv.org/abs/2603.00466 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00466.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Boming Tan, Xiangdong Zhang, Ning Liao, Yuqing Zhang, Shaofeng Zhang, Xue Yang, Qi Fan, Yanyong Zhang</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> Despite impressive progress in video generation, existing models remain limited to surface-level plausibility, lacking a coherent and unified understanding of the world. Prior approaches typically incorporate only a single form of world-related knowledge or rely on rigid alignment strategies to introduce additional knowledge. However, aligning the single world knowledge is insufficient to constitute a world model that requires jointly modeling multiple heterogeneous dimensions (e.g., physical commonsense, 3D and temporal consistency). To address this limitation, we introduce DreamWorld, a unified framework that integrates complementary world knowledge into video generators via a Joint World Modeling Paradigm, jointly predicting video pixels and features from foundation models to capture temporal dynamics, spatial geometry, and semantic consistency. However, naively optimizing these heterogeneous objectives can lead to visual instability and temporal flickering. To mitigate this issue, we propose Consistent Constraint Annealing (CCA) to progressively regulate world-level constraints during training, and Multi-Source Inner-Guidance to enforce learned world priors at inference. Extensive evaluations show that DreamWorld improves world consistency, outperforming Wan2.1 by 2.26 points on VBench. Code will be made publicly available at https://github.com/ABU121111/DreamWorld{mypink{Github}}.</p> - https://arxiv.org/abs/2603.00466 - Sat, 28 Feb 2026 05:02:39 +0000 + Hindsight Credit Assignment for Long-Horizon LLM Agents + https://arxiv.org/abs/2603.08754 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08754.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hui-Ze Tan, Xiao-Wen Yang, Hao Chen, Jie-Jing Shao, Yi Wen, Yuteng Shen, Weihong Luo, Xiku Du, Lan-Zhe Guo, Yu-Feng Li</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Large Language Model (LLM) agents often face significant credit assignment challenges in long-horizon, multi-step tasks due to sparse rewards. Existing value-free methods, such as Group Relative Policy Optimization (GRPO), encounter two fundamental bottlenecks: inaccurate step-level Q-value estimation and misaligned value baselines for intermediate states. To address these limitations, we introduce HCAPO, the first framework to integrate hindsight credit assignment into LLM agents. HCAPO leverages the LLM itself as a post-hoc critic to refine step-level Q-values through hindsight reasoning. Furthermore, HCAPO's multi-scale advantage mechanism effectively supplements the inaccurate value baselines at critical decision states. Evaluations across three challenging benchmarks, including WebShop and ALFWorld, demonstrate that HCAPO consistently outperforms state-of-the-art RL methods. Notably, HCAPO achieves a 7.7% improvement in success rate on WebShop and a 13.8% on ALFWorld over GRPO using the Qwen2.5-7B-Instruct model. These results indicate that HCAPO significantly enhances exploration efficiency, promotes concise decision-making, and ensures scalability in complex, long-horizon tasks.</p> + https://arxiv.org/abs/2603.08754 + Sat, 07 Mar 2026 06:05:20 +0000 - Spectral Condition for μP under Width-Depth Scaling - https://arxiv.org/abs/2603.00541 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00541.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chenyu Zheng, Rongzhen Wang, Xinyu Zhang, Chongxuan Li</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> Generative foundation models are increasingly scaled in both width and depth, posing significant challenges for stable feature learning and reliable hyperparameter (HP) transfer across model sizes. While maximal update parameterization (μP) has provided a principled solution to both problems for width scaling, existing extensions to the joint width-depth scaling regime remain fragmented, architecture- and optimizer-specific, and often rely on technically involved theories. In this work, we develop a simple and unified spectral framework for μP under joint width-depth scaling. Considering residual networks of varying block depths, we first introduce a spectral μP condition that precisely characterizes how the norms of weights and their per-step updates should scale with width and depth, unifying previously disparate μP formulations as special cases. Building on this condition, we then derive a general recipe for implementing μP across a broad class of optimizers by mapping the spectral constraints to concrete HP parameterizations. This approach not only recovers existing μP formulations (e.g., for SGD and AdamW) but also naturally extends to a wider range of optimizers. Finally, experiments on GPT-2 style language models demonstrate that the proposed spectral μP condition preserves stable feature learning and enables robust HP transfer under width-depth scaling.</p> - https://arxiv.org/abs/2603.00541 - Sat, 28 Feb 2026 08:38:50 +0000 + Unlocking Data Value in Finance: A Study on Distillation and Difficulty-Aware Training + https://arxiv.org/abs/2603.07223 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07223.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chuxue Cao, Honglin Lin, Zhanping Zhong, Xin Gao, Mengzhang Cai, Conghui He, Sirui Han, Lijun Wu</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Large Language Models (LLMs) have demonstrated strong general capabilities, yet their deployment in finance remains challenging due to dense domain-specific terminology, stringent numerical reasoning requirements, and low tolerance for factual errors. We conduct a controlled empirical study showing that in specialized vertical domains, performance is largely determined by the quality and difficulty/verifiability profile of post-training data. We introduce ODA-Fin-SFT-318k, constructed via multi-stage distillation and verification to produce high-quality Chain-of-Thought supervision, and ODA-Fin-RL-12k, curated for hard-but-verifiable tasks that balance reward precision and task diversity. Using standard SFT and RL pipelines, we show that high-quality CoT distillation establishes a robust foundation during SFT, while difficulty- and verifiability-aware sampling improves RL generalization. Evaluated on nine benchmarks spanning general financial tasks, sentiment analysis, and numerical reasoning, our ODA-Fin-RL-8B consistently surpasses open-source state-of-the-art (SOTA) financial LLMs of comparable size. We release our ODA-Fin-SFT-318k and ODA-Fin-RL-12k datasets, along with trained models to advance data-centric financial AI research.</p> + https://arxiv.org/abs/2603.07223 + Sat, 07 Mar 2026 14:07:37 +0000 - CMI-RewardBench: Evaluating Music Reward Models with Compositional Multimodal Instruction - https://arxiv.org/abs/2603.00610 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00610.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yinghao Ma, Haiwen Xia, Hewei Gao, Weixiong Chen, Yuxin Ye, Yuchen Yang, Sungkyun Chang, Mingshuo Ding, Yizhi Li, Ruibin Yuan, Simon Dixon, Emmanouil Benetos</p><p><b>Upvotes:</b> 32</p><p><b>Summary:</b> While music generation models have evolved to handle complex multimodal inputs mixing text, lyrics, and reference audio, evaluation mechanisms have lagged behind. In this paper, we bridge this critical gap by establishing a comprehensive ecosystem for music reward modeling under Compositional Multimodal Instruction (CMI), where the generated music may be conditioned on text descriptions, lyrics, and audio prompts. We first introduce CMI-Pref-Pseudo, a large-scale preference dataset comprising 110k pseudo-labeled samples, and CMI-Pref, a high-quality, human-annotated corpus tailored for fine-grained alignment tasks. To unify the evaluation landscape, we propose CMI-RewardBench, a unified benchmark that evaluates music reward models on heterogeneous samples across musicality, text-music alignment, and compositional instruction alignment. Leveraging these resources, we develop CMI reward models (CMI-RMs), a parameter-efficient reward model family capable of processing heterogeneous inputs. We evaluate their correlation with human judgments scores on musicality and alignment on CMI-Pref along with previous datasets. Further experiments demonstrate that CMI-RM not only correlates strongly with human judgments, but also enables effective inference-time scaling via top-k filtering. The necessary training data, benchmarks, and reward models are publicly available.</p> - https://arxiv.org/abs/2603.00610 - Sat, 28 Feb 2026 12:10:58 +0000 + AutoResearch-RL: Perpetual Self-Evaluating Reinforcement Learning Agents for Autonomous Neural Architecture Discovery + https://arxiv.org/abs/2603.07300 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07300.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Nilesh Jain, Rohit Yadav, Sagar Kotian, Claude AI</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> We present AutoResearch-RL, a framework in which a reinforcement learning agent conducts open-ended neural architecture and hyperparameter research without human supervision, running perpetually until a termination oracle signals convergence or resource exhaustion. At each step the agent proposes a code modification to a target training script, executes it under a fixed wall clock time budget, observes a scalar reward derived from validation bits-per-byte (val-bpb), and updates its policy via Proximal Policy Optimisation (PPO). The key design insight is the separation of three concerns: (i) a frozen environment (data pipeline, evaluation protocol, and constants) that guarantees fair cross-experiment comparison; (ii) a mutable target file (train.py) that represents the agent's editable state; and (iii) a meta-learner (the RL agent itself) that accumulates a growing trajectory of experiment outcomes and uses them to inform subsequent proposals. We formalise this as a Markov Decision Process, derive convergence guarantees under mild assumptions, and demonstrate empirically on a single GPU nanochat pretraining benchmark that AutoResearch-RL discovers configurations that match or exceed hand-tuned baselines after approximately 300 overnight iterations, with no human in the loop.</p> + https://arxiv.org/abs/2603.07300 + Sat, 07 Mar 2026 17:49:44 +0000 - InfoPO: Information-Driven Policy Optimization for User-Centric Agents - https://arxiv.org/abs/2603.00656 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00656.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Fanqi Kong, Jiayi Zhang, Mingyi Deng, Chenglin Wu, Yuyu Luo, Bang Liu</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Real-world user requests to LLM agents are often underspecified. Agents must interact to acquire missing information and make correct downstream decisions. However, current multi-turn GRPO-based methods often rely on trajectory-level reward computation, which leads to credit assignment problems and insufficient advantage signals within rollout groups. A feasible approach is to identify valuable interaction turns at a fine granularity to drive more targeted learning. To address this, we introduce InfoPO (Information-Driven Policy Optimization), which frames multi-turn interaction as a process of active uncertainty reduction and computes an information-gain reward that credits turns whose feedback measurably changes the agent's subsequent action distribution compared to a masked-feedback counterfactual. It then combines this signal with task outcomes via an adaptive variance-gated fusion to identify information importance while maintaining task-oriented goal direction. Across diverse tasks, including intent clarification, collaborative coding, and tool-augmented decision making, InfoPO consistently outperforms prompting and multi-turn RL baselines. It also demonstrates robustness under user simulator shifts and generalizes effectively to environment-interactive tasks. Overall, InfoPO provides a principled and scalable mechanism for optimizing complex agent-user collaboration. Code is available at https://github.com/kfq20/InfoPO.</p> - https://arxiv.org/abs/2603.00656 - Sat, 28 Feb 2026 13:58:14 +0000 + Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams + https://arxiv.org/abs/2603.07392 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07392.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiyeon Kim, Hyunji Lee, Dylan Zhou, Sue Hyun Park, Seunghyun Yoon, Trung Bui, Franck Dernoncourt, Sungmin Cha, Minjoon Seo</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> LLMs operating in dynamic real-world contexts often encounter knowledge that evolves continuously or emerges incrementally. To remain accurate and effective, models must adapt to newly arriving information on the fly. We introduce Online Adaptation to Continual Knowledge Streams(OAKS) to evaluate this capability, establishing a benchmark for online adaptation over streaming, continually updating knowledge. Specifically, the benchmark is structured as a sequence of fine-grained context chunks where facts change dynamically across time intervals. OAKS comprises two datasets: OAKS-BABI and OAKS-Novel, where individual facts evolve multiple times across context chunks. These datasets include dense annotations to measure whether models track changes accurately. Evaluating 14 models with varied inference approaches, we observe significant limitations in current methodologies. Both state-of-the-art models and agentic memory systems fail to adapt robustly on OAKS, demonstrating delays in state-tracking and susceptibility to distraction within streaming environments.</p> + https://arxiv.org/abs/2603.07392 + Sun, 08 Mar 2026 00:38:39 +0000 - STMI: Segmentation-Guided Token Modulation with Cross-Modal Hypergraph Interaction for Multi-Modal Object Re-Identification - https://arxiv.org/abs/2603.00695 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00695.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xingguo Xu, Zhanyu Liu, Weixiang Zhou, Yuansheng Gao, Junjie Cao, Yuhao Wang, Jixiang Luo, Dell Zhang</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Multi-modal object Re-Identification (ReID) aims to exploit complementary information from different modalities to retrieve specific objects. However, existing methods often rely on hard token filtering or simple fusion strategies, which can lead to the loss of discriminative cues and increased background interference. To address these challenges, we propose STMI, a novel multi-modal learning framework consisting of three key components: (1) Segmentation-Guided Feature Modulation (SFM) module leverages SAM-generated masks to enhance foreground representations and suppress background noise through learnable attention modulation; (2) Semantic Token Reallocation (STR) module employs learnable query tokens and an adaptive reallocation mechanism to extract compact and informative representations without discarding any tokens; (3) Cross-Modal Hypergraph Interaction (CHI) module constructs a unified hypergraph across modalities to capture high-order semantic relationships. Extensive experiments on public benchmarks (i.e., RGBNT201, RGBNT100, and MSVR310) demonstrate the effectiveness and robustness of our proposed STMI framework in multi-modal ReID scenarios.</p> - https://arxiv.org/abs/2603.00695 - Sat, 28 Feb 2026 15:07:10 +0000 + Accent Vector: Controllable Accent Manipulation for Multilingual TTS Without Accented Data + https://arxiv.org/abs/2603.07534 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07534.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Thanathai Lertpetchpun, Thanapat Trachu, Jihwan Lee, Tiantian Feng, Dani Byrd, Shrikanth Narayanan</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Accent is an integral part of society, reflecting multiculturalism and shaping how individuals express identity. The majority of English speakers are non-native (L2) speakers, yet current Text-To-Speech (TTS) systems primarily model American-accented English due limited accented data. We propose Accent Vector, a controllable representation that enables accent manipulation in multilingual TTS without requiring accented training data. Accent Vector is derived by fine-tuning a TTS system on native speech of a different language (i.e. non-English) and computing task vectors capturing accent characteristics (i.e. in English). By scaling and interpolating the vector, we achieve fine-grained control over accent strength and generate mixed-accent speech. In addition, it generalizes beyond English, enabling accent control across multiple languages. Objective and human evaluations confirm the effectiveness of Accent Vector for fine-grained and compositional accent control.</p> + https://arxiv.org/abs/2603.07534 + Sun, 08 Mar 2026 08:48:42 +0000 - Qwen3-Coder-Next Technical Report - https://arxiv.org/abs/2603.00729 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00729.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ruisheng Cao, Mouxiang Chen, Jiawei Chen, Zeyu Cui, Yunlong Feng, Binyuan Hui, Yuheng Jing, Kaixin Li, Mingze Li, Junyang Lin, Zeyao Ma, Kashun Shum, Xuwu Wang, Jinxi Wei, Jiaxi Yang, Jiajun Zhang, Lei Zhang, Zongmeng Zhang, Wenting Zhao, Fan Zhou</p><p><b>Upvotes:</b> 39</p><p><b>Summary:</b> We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. Qwen3-Coder-Next is an 80-billion-parameter model that activates only 3 billion parameters during inference, enabling strong coding capability with efficient inference. In this work, we explore how far strong training recipes can push the capability limits of models with small parameter footprints. To achieve this, we perform agentic training through large-scale synthesis of verifiable coding tasks paired with executable environments, allowing learning directly from environment feedback via mid-training and reinforcement learning. Across agent-centric benchmarks including SWE-Bench and Terminal-Bench, Qwen3-Coder-Next achieves competitive performance relative to its active parameter count. We release both base and instruction-tuned open-weight versions to support research and real-world coding agent development.</p> - https://arxiv.org/abs/2603.00729 - Sat, 28 Feb 2026 16:25:04 +0000 + PureCC: Pure Learning for Text-to-Image Concept Customization + https://arxiv.org/abs/2603.07561 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07561.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhichao Liao, Xiaole Xian, Qingyu Li, Wenyu Qin, Meng Wang, Weicheng Xie, Siyang Song, Pingfa Feng, Long Zeng, Liang Pan</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Existing concept customization methods have achieved remarkable outcomes in high-fidelity and multi-concept customization. However, they often neglect the influence on the original model's behavior and capabilities when learning new personalized concepts. To address this issue, we propose PureCC. PureCC introduces a novel decoupled learning objective for concept customization, which combines the implicit guidance of the target concept with the original conditional prediction. This separated form enables PureCC to substantially focus on the original model during training. Moreover, based on this objective, PureCC designs a dual-branch training pipeline that includes a frozen extractor providing purified target concept representations as implicit guidance and a trainable flow model producing the original conditional prediction, jointly achieving pure learning for personalized concepts. Furthermore, PureCC introduces a novel adaptive guidance scale λ^star to dynamically adjust the guidance strength of the target concept, balancing customization fidelity and model preservation. Extensive experiments show that PureCC achieves state-of-the-art performance in preserving the original behavior and capabilities while enabling high-fidelity concept customization. The code is available at https://github.com/lzc-sg/PureCC.</p> + https://arxiv.org/abs/2603.07561 + Sun, 08 Mar 2026 09:50:20 +0000 - CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning - https://arxiv.org/abs/2603.00889 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00889.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xinyu Zhu, Yihao Feng, Yanchao Sun, Xianzhi Du, Pingzhi Li, Olli Saarikivi, Yun Zhu, Yu Meng</p><p><b>Upvotes:</b> 43</p><p><b>Summary:</b> Large Language Models (LLMs) have recently exhibited remarkable reasoning capabilities, largely enabled by supervised fine-tuning (SFT)- and reinforcement learning (RL)-based post-training on high-quality reasoning data. However, reproducing and extending these capabilities in open and scalable settings is hindered by three fundamental data-centric challenges: (1) the cold-start problem, arising from the lack of seed datasets with detailed, long Chain-of-Thought (CoT) trajectories needed to initialize reasoning policies; (2) limited domain coverage, as most existing open-source reasoning datasets are concentrated in mathematics, with limited coverage of broader scientific disciplines; and (3) the annotation bottleneck, where the difficulty of frontier-level reasoning tasks makes reliable human annotation prohibitively expensive or infeasible. To address these challenges, we introduce CHIMERA, a compact synthetic reasoning dataset comprising 9K samples for generalizable cross-domain reasoning. CHIMERA is constructed with three key properties: (1) it provides rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models; (2) it has broad and structured coverage, spanning 8 major scientific disciplines and over 1K fine-grained topics organized via a model-generated hierarchical taxonomy; and (3) it employs a fully automated, scalable evaluation pipeline that uses strong reasoning models to cross-validate both problem validity and answer correctness. We use CHIMERA to post-train a 4B Qwen3 model. Despite the dataset's modest size, the resulting model achieves strong performance on a suite of challenging reasoning benchmarks, including GPQA-Diamond, AIME 24/25/26, HMMT 25, and Humanity's Last Exam, approaching or matching the reasoning performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.</p> - https://arxiv.org/abs/2603.00889 - Sun, 01 Mar 2026 03:23:41 +0000 + Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence + https://arxiv.org/abs/2603.07660 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07660.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuanyuan Gao, Hao Li, Yifei Liu, Xinhao Ji, Yuning Gong, Yuanjun Liao, Fangfu Liu, Manyuan Zhang, Yuchen Yang, Dan Xu, Xue Yang, Huaxi Huang, Hongjie Zhang, Ziwei Liu, Xiao Sun, Dingwen Zhang, Zhihang Zhong</p><p><b>Upvotes:</b> 77</p><p><b>Summary:</b> The pursuit of spatial intelligence fundamentally relies on access to large-scale, fine-grained 3D data. However, existing approaches predominantly construct spatial understanding benchmarks by generating question-answer (QA) pairs from a limited number of manually annotated datasets, rather than systematically annotating new large-scale 3D scenes from raw web data. As a result, their scalability is severely constrained, and model performance is further hindered by domain gaps inherent in these narrowly curated datasets. In this work, we propose Holi-Spatial, the first fully automated, large-scale, spatially-aware multimodal dataset, constructed from raw video inputs without human intervention, using the proposed data curation pipeline. Holi-Spatial supports multi-level spatial supervision, ranging from geometrically accurate 3D Gaussian Splatting (3DGS) reconstructions with rendered depth maps to object-level and relational semantic annotations, together with corresponding spatial Question-Answer (QA) pairs. Following a principled and systematic pipeline, we further construct Holi-Spatial-4M, the first large-scale, high-quality 3D semantic dataset, containing 12K optimized 3DGS scenes, 1.3M 2D masks, 320K 3D bounding boxes, 320K instance captions, 1.2M 3D grounding instances, and 1.2M spatial QA pairs spanning diverse geometric, relational, and semantic reasoning tasks. Holi-Spatial demonstrates exceptional performance in data curation quality, significantly outperforming existing feed-forward and per-scene optimized methods on datasets such as ScanNet, ScanNet++, and DL3DV. Furthermore, fine-tuning Vision-Language Models (VLMs) on spatial reasoning tasks using this dataset has also led to substantial improvements in model performance.</p> + https://arxiv.org/abs/2603.07660 + Sun, 08 Mar 2026 14:49:20 +0000 - VGGT-Det: Mining VGGT Internal Priors for Sensor-Geometry-Free Multi-View Indoor 3D Object Detection - https://arxiv.org/abs/2603.00912 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.00912.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yang Cao, Feize Wu, Dave Zhenyu Chen, Yingji Zhong, Lanqing Hong, Dan Xu</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> Current multi-view indoor 3D object detectors rely on sensor geometry that is costly to obtain (i.e., precisely calibrated multi-view camera poses) to fuse multi-view information into a global scene representation, limiting deployment in real-world scenes. We target a more practical setting: Sensor-Geometry-Free (SG-Free) multi-view indoor 3D object detection, where there are no sensor-provided geometric inputs (multi-view poses or depth). Recent Visual Geometry Grounded Transformer (VGGT) shows that strong 3D cues can be inferred directly from images. Building on this insight, we present VGGT-Det, the first framework tailored for SG-Free multi-view indoor 3D object detection. Rather than merely consuming VGGT predictions, our method integrates VGGT encoder into a transformer-based pipeline. To effectively leverage both the semantic and geometric priors from inside VGGT, we introduce two novel key components: (i) Attention-Guided Query Generation (AG): exploits VGGT attention maps as semantic priors to initialize object queries, improving localization by focusing on object regions while preserving global spatial structure; (ii) Query-Driven Feature Aggregation (QD): a learnable See-Query interacts with object queries to 'see' what they need, and then dynamically aggregates multi-level geometric features across VGGT layers that progressively lift 2D features into 3D. Experiments show that VGGT-Det significantly surpasses the best-performing method in the SG-Free setting by 4.4 and 8.6 mAP@0.25 on ScanNet and ARKitScenes, respectively. Ablation study shows that VGGT's internally learned semantic and geometric priors can be effectively leveraged by our AG and QD.</p> - https://arxiv.org/abs/2603.00912 - Sun, 01 Mar 2026 04:25:52 +0000 + TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward + https://arxiv.org/abs/2603.07700 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07700.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yihong Luo, Tianyang Hu, Weijian Luo, Jing Tang</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> While few-step generative models have enabled powerful image and video generation at significantly lower cost, generic reinforcement learning (RL) paradigms for few-step models remain an unsolved problem. Existing RL approaches for few-step diffusion models strongly rely on back-propagating through differentiable reward models, thereby excluding the majority of important real-world reward signals, e.g., non-differentiable rewards such as humans' binary likeness, object counts, etc. To properly incorporate non-differentiable rewards to improve few-step generative models, we introduce TDM-R1, a novel reinforcement learning paradigm built upon a leading few-step model, Trajectory Distribution Matching (TDM). TDM-R1 decouples the learning process into surrogate reward learning and generator learning. Furthermore, we developed practical methods to obtain per-step reward signals along the deterministic generation trajectory of TDM, resulting in a unified RL post-training method that significantly improves few-step models' ability with generic rewards. We conduct extensive experiments ranging from text-rendering, visual quality, and preference alignment. All results demonstrate that TDM-R1 is a powerful reinforcement learning paradigm for few-step text-to-image models, achieving state-of-the-art reinforcement learning performances on both in-domain and out-of-domain metrics. Furthermore, TDM-R1 also scales effectively to the recent strong Z-Image model, consistently outperforming both its 100-NFE and few-step variants with only 4 NFEs. Project page: https://github.com/Luo-Yihong/TDM-R1</p> + https://arxiv.org/abs/2603.07700 + Sun, 08 Mar 2026 15:57:10 +0000 - LLaDA-o: An Effective and Length-Adaptive Omni Diffusion Model - https://arxiv.org/abs/2603.01068 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01068.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zebin You, Xiaolu Zhang, Jun Zhou, Chongxuan Li, Ji-Rong Wen</p><p><b>Upvotes:</b> 19</p><p><b>Summary:</b> We present LLaDA-o, an effective and length-adaptive omni diffusion model for multimodal understanding and generation. LLaDA-o is built on a Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text understanding and continuous diffusion for visual generation, while coupling them through a shared, simple, and efficient attention backbone that reduces redundant computation for fixed conditions. Building on MoD, we further introduce a data-centric length adaptation strategy that enables flexible-length decoding in multimodal settings without architectural changes. Extensive experiments show that LLaDA-o achieves state-of-the-art performance among omni-diffusion models on multimodal understanding and generation benchmarks, and reaches 87.04 on DPG-Bench for text-to-image generation, supporting the effectiveness of unified omni diffusion modeling. Code is available at https://github.com/ML-GSAI/LLaDA-o.</p> - https://arxiv.org/abs/2603.01068 - Sun, 01 Mar 2026 12:05:06 +0000 + VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning? + https://arxiv.org/abs/2603.07888 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07888.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Minkyu Kim, Sangheon Lee, Dongmin Park</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> The ability to distinguish subtle differences between visually similar images is essential for diverse domains such as industrial anomaly detection, medical imaging, and aerial surveillance. While comparative reasoning benchmarks for vision-language models (VLMs) have recently emerged, they primarily focus on images with large, salient differences and fail to capture the nuanced reasoning required for real-world applications. In this work, we introduce VLM-SubtleBench, a benchmark designed to evaluate VLMs on subtle comparative reasoning. Our benchmark covers ten difference types - Attribute, State, Emotion, Temporal, Spatial, Existence, Quantity, Quality, Viewpoint, and Action - and curate paired question-image sets reflecting these fine-grained variations. Unlike prior benchmarks restricted to natural image datasets, our benchmark spans diverse domains, including industrial, aerial, and medical imagery. Through extensive evaluation of both proprietary and open-source VLMs, we reveal systematic gaps between model and human performance across difference types and domains, and provide controlled analyses highlighting where VLMs' reasoning sharply deteriorates. Together, our benchmark and findings establish a foundation for advancing VLMs toward human-level comparative reasoning.</p> + https://arxiv.org/abs/2603.07888 + Mon, 09 Mar 2026 02:01:02 +0000 - Unified Vision-Language Modeling via Concept Space Alignment - https://arxiv.org/abs/2603.01096 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01096.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yifu Qiu, Paul-Ambroise Duquenne, Holger Schwenk</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> We introduce V-SONAR, a vision-language embedding space extended from the text-only embedding space SONAR (Omnilingual Embeddings Team et al., 2026), which supports 1500 text languages and 177 speech languages. To construct V-SONAR, we propose a post-hoc alignment pipeline that maps the representations of an existing vision encoder into the SONAR space. We thoroughly evaluate V-SONAR and show that its embeddings achieve competitive performance on text-to-video retrieval. Equipped with the OMNISONAR text decoder, V-SONAR further surpasses state-of-the-art vision-language models on video captioning tasks, including DREAM-1K (BLEU 23.9 vs. 19.6) and PE-VIDEO (BLEU 39.0 vs. 30.0). Leveraging V-SONAR, we first demonstrate that the Large Concept Model (LCM; LCM team et al. 2024) operating in SONAR and trained with English text only, can perform both single- and multi-visual concept understanding in a zero-shot manner. Finally, we introduce V-LCM, which extends the LCM with vision-language instruction tuning. V-LCM encodes vision and language inputs into an unified sequence of latent embeddings via V-SONAR and SONAR, and it is trained with the same latent diffusion objective for next-embedding prediction as in LCM's text-only pre-training. Experiments on a large-scale multilingual and -modal instruction-tuning data mixture highlight the potential of V-LCM: V-LCM matches state-of-the-art vision-language models on tasks covering image/video captioning and question answering, while significantly outperforming them across 61 rich- to low-resource languages out of all 62 tested languages.</p> - https://arxiv.org/abs/2603.01096 - Sun, 01 Mar 2026 13:22:44 +0000 + \$OneMillion-Bench: How Far are Language Agents from Human Experts? + https://arxiv.org/abs/2603.07980 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.07980.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qianyu Yang, Yang Liu, Jiaqi Li, Jun Bai, Hao Chen, Kaiyuan Chen, Tiliang Duan, Jiayun Dong, Xiaobo Hu, Zixia Jia, Yang Liu, Tao Peng, Yixin Ren, Ran Tian, Zaiyuan Wang, Yanglihong Xiao, Gang Yao, Lingyue Yin, Ge Zhang, Chun Zhang, Jianpeng Jiao, Zilong Zheng, Yuan Gong</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> As language models (LMs) evolve from chat assistants to long-horizon agents capable of multi-step reasoning and tool use, existing benchmarks remain largely confined to structured or exam-style tasks that fall short of real-world professional demands. To this end, we introduce \OneMillion-Bench OneMillion-Bench, a benchmark of 400 expert-curated tasks spanning Law, Finance, Industry, Healthcare, and Natural Science, built to evaluate agents across economically consequential scenarios. Unlike prior work, the benchmark requires retrieving authoritative sources, resolving conflicting evidence, applying domain-specific rules, and making constraint decisions, where correctness depends as much on the reasoning process as the final answer. We adopt a rubric-based evaluation protocol scoring factual accuracy, logical coherence, practical feasibility, and professional compliance, focused on expert-level problems to ensure meaningful differentiation across agents. Together, \$OneMillion-Bench provides a unified testbed for assessing agentic reliability, professional depth, and practical readiness in domain-intensive scenarios.</p> + https://arxiv.org/abs/2603.07980 + Mon, 09 Mar 2026 05:32:42 +0000 - Learn Hard Problems During RL with Reference Guided Fine-tuning - https://arxiv.org/abs/2603.01223 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01223.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yangzhen Wu, Shanda Li, Zixin Wen, Xin Zhou, Ameet Talwalkar, Yiming Yang, Wenhao Huang, Tianle Cai</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Reinforcement learning (RL) for mathematical reasoning can suffer from reward sparsity: for challenging problems, LLM fails to sample any correct trajectories, preventing RL from receiving meaningful positive feedback. At the same time, there often exist human-written reference solutions along with the problem (e.g., problems from AoPS), but directly fine-tuning on these solutions offers no benefit because models often cannot imitate human proofs that lie outside their own reasoning distribution. We introduce Reference-Guided Fine-Tuning (ReGFT), a simple and effective method that utilizes human-written reference solutions to synthesize positive trajectories on hard problems and train on them before RL. For each problem, we provide the model with a partial reference solution and let it generate its own reasoning trace, ensuring the resulting trajectories remain in the model's reasoning space while still benefiting from reference guidance. Fine-tuning on these reference-guided trajectories increases the number of solvable problems and produces a checkpoint that receives more positive rewards during RL. Across three benchmarks (AIME24, AIME25, BeyondAIME), ReGFT consistently improves supervised accuracy, accelerates DAPO training, and raises the final performance plateau of RL. Our results show that ReGFT effectively overcomes reward sparsity and unlocks stronger RL-based mathematical reasoning.</p> - https://arxiv.org/abs/2603.01223 - Sun, 01 Mar 2026 18:41:28 +0000 + PIRA-Bench: A Transition from Reactive GUI Agents to GUI-based Proactive Intent Recommendation Agents + https://arxiv.org/abs/2603.08013 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08013.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuxiang Chai, Shunye Tang, Han Xiao, Rui Liu, Hongsheng Li</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Current Graphical User Interface (GUI) agents operate primarily under a reactive paradigm: a user must provide an explicit instruction for the agent to execute a task. However, an intelligent AI assistant should be proactive, which is capable of anticipating user intentions directly from continuous visual inputs, such as mobile or desktop screenshots, and offering timely recommendations without explicit user prompting. Transitioning to this proactive paradigm presents significant challenges. Real-world screen activity is rarely linear; it consists of long-horizon trajectories fraught with noisy browsing, meaningless actions, and multithreaded task-switching. To address this gap, we introduce PIRA-Bench (Proactive Intent Recommendation Agent Benchmark), a novel benchmark for evaluating multimodal large language models (MLLMs) on continuous, weakly-supervised visual inputs. Unlike reactive datasets, PIRA-Bench features complex trajectories with multiple interleaved intents and noisy segments with various user profile contexts, challenging agents to detect actionable events while fitting to user preferences. Furthermore, we propose the PIRF baseline, a memory-aware, state-tracking framework that empowers general MLLMs to manage multiple task threads and handle misleading visual inputs. PIRA-Bench serves as an initial step toward robust and proactive GUI-based personal assistants.</p> + https://arxiv.org/abs/2603.08013 + Mon, 09 Mar 2026 06:41:32 +0000 - AgilePruner: An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models - https://arxiv.org/abs/2603.01236 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01236.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Changwoo Baek, Jouwon Song, Sohyeon Kim, Kyeongbo Kong</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Large Vision-Language Models (LVLMs) have adopted visual token pruning strategies to mitigate substantial computational overhead incurred by extensive visual token sequences. While prior works primarily focus on either attention-based or diversity-based pruning methods, in-depth analysis of these approaches' characteristics and limitations remains largely unexplored. In this work, we conduct thorough empirical analysis using effective rank (erank) as a measure of feature diversity and attention score entropy to investigate visual token processing mechanisms and analyze the strengths and weaknesses of each approach. Our analysis reveals two insights: (1) Our erank-based quantitative analysis shows that many diversity-oriented pruning methods preserve substantially less feature diversity than intended; moreover, analysis using the CHAIR dataset reveals that the diversity they do retain is closely tied to increased hallucination frequency compared to attention-based pruning. (2) We further observe that attention-based approaches are more effective on simple images where visual evidence is concentrated, while diversity-based methods better handle complex images with distributed features. Building on these empirical insights, we show that incorporating image-aware adjustments into existing hybrid pruning strategies consistently improves their performance. We also provide a minimal instantiation of our empirical findings through a simple adaptive pruning mechanism, which achieves strong and reliable performance across standard benchmarks as well as hallucination-specific evaluations. Our project page available at https://cvsp-lab.github.io/AgilePruner.</p> - https://arxiv.org/abs/2603.01236 - Sun, 01 Mar 2026 19:14:39 +0000 + In-Context Reinforcement Learning for Tool Use in Large Language Models + https://arxiv.org/abs/2603.08068 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08068.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yaoqi Ye, Yiran Zhao, Keyu Duan, Zeyu Zheng, Kenji Kawaguchi, Cihang Xie, Michael Qizhe Shieh</p><p><b>Upvotes:</b> 28</p><p><b>Summary:</b> While large language models (LLMs) exhibit strong reasoning abilities, their performance on complex tasks is often constrained by the limitations of their internal knowledge. A compelling approach to overcome this challenge is to augment these models with external tools -- such as Python interpreters for mathematical computations or search engines for retrieving factual information. However, enabling models to use these tools effectively remains a significant challenge. Existing methods typically rely on cold-start pipelines that begin with supervised fine-tuning (SFT), followed by reinforcement learning (RL). These approaches often require substantial amounts of labeled data for SFT, which is expensive to annotate or synthesize. In this work, we propose In-Context Reinforcement Learning (ICRL), an RL-only framework that eliminates the need for SFT by leveraging few-shot prompting during the rollout stage of RL. Specifically, ICRL introduces in-context examples within the rollout prompts to teach the model how to invoke external tools. Furthermore, as training progresses, the number of in-context examples is gradually reduced, eventually reaching a zero-shot setting where the model learns to call tools independently. We conduct extensive experiments across a range of reasoning and tool-use benchmarks. Results show that ICRL achieves state-of-the-art performance, demonstrating its effectiveness as a scalable, data-efficient alternative to traditional SFT-based pipelines.</p> + https://arxiv.org/abs/2603.08068 + Mon, 09 Mar 2026 08:06:18 +0000 - Spectral Attention Steering for Prompt Highlighting - https://arxiv.org/abs/2603.01281 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01281.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Weixian Waylon Li, Yuchen Niu, Yongxin Yang, Keshuang Li, Tiejun Ma, Shay B. Cohen</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Attention steering is an important technique for controlling model focus, enabling capabilities such as prompt highlighting, where the model prioritises user-specified text. However, existing attention steering methods require explicit storage of the full attention matrix, making them incompatible with memory-efficient implementations like FlashAttention. We introduce Spectral Editing Key Amplification (SEKA), a training-free steering method that tackles this by directly editing key embeddings before attention computation. SEKA uses spectral decomposition to steer key embeddings towards latent directions that amplify attention scores for certain tokens. We extend this to Adaptive SEKA (AdaSEKA), a query-adaptive variant that uses a training-free routing mechanism to dynamically combine multiple expert subspaces based on the prompt's semantic intent. Our experiments show both methods significantly outperform strong baselines on standard steering benchmarks while adding much lower latency and memory overhead, in compatibility with optimised attention.</p> - https://arxiv.org/abs/2603.01281 - Sun, 01 Mar 2026 21:35:09 +0000 + Concept-Guided Fine-Tuning: Steering ViTs away from Spurious Correlations to Improve Robustness + https://arxiv.org/abs/2603.08309 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08309.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yehonatan Elisha, Oren Barkan, Noam Koenigstein</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Vision Transformers (ViTs) often degrade under distribution shifts because they rely on spurious correlations, such as background cues, rather than semantically meaningful features. Existing regularization methods, typically relying on simple foreground-background masks, which fail to capture the fine-grained semantic concepts that define an object (e.g., ``long beak'' and ``wings'' for a ``bird''). As a result, these methods provide limited robustness to distribution shifts. To address this limitation, we introduce a novel finetuning framework that steers model reasoning toward concept-level semantics. Our approach optimizes the model's internal relevance maps to align with spatially grounded concept masks. These masks are generated automatically, without manual annotation: class-relevant concepts are first proposed using an LLM-based, label-free method, and then segmented using a VLM. The finetuning objective aligns relevance with these concept regions while simultaneously suppressing focus on spurious background areas. Notably, this process requires only a minimal set of images and uses half of the dataset classes. Extensive experiments on five out-of-distribution benchmarks demonstrate that our method improves robustness across multiple ViT-based models. Furthermore, we show that the resulting relevance maps exhibit stronger alignment with semantic object parts, offering a scalable path toward more robust and interpretable vision models. Finally, we confirm that concept-guided masks provide more effective supervision for model robustness than conventional segmentation maps, supporting our central hypothesis.</p> + https://arxiv.org/abs/2603.08309 + Mon, 09 Mar 2026 12:31:14 +0000 - When Does RL Help Medical VLMs? Disentangling Vision, SFT, and RL Gains - https://arxiv.org/abs/2603.01301 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01301.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ahmadreza Jeddi, Kimia Shaban, Negin Baghbanzadeh, Natasha Sharan, Abhishek Moturu, Elham Dolatabadi, Babak Taati</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Reinforcement learning (RL) is increasingly used to post-train medical Vision-Language Models (VLMs), yet it remains unclear whether RL improves medical visual reasoning or mainly sharpens behaviors already induced by supervised fine-tuning (SFT). We present a controlled study that disentangles these effects along three axes: vision, SFT, and RL. Using MedMNIST as a multi-modality testbed, we probe visual perception by benchmarking VLM vision towers against vision-only baselines, quantify reasoning support and sampling efficiency via Accuracy@1 versus Pass@K, and evaluate when RL closes the support gap and how gains transfer across modalities. We find that RL is most effective when the model already has non-trivial support (high Pass@K): it primarily sharpens the output distribution, improving Acc@1 and sampling efficiency, while SFT expands support and makes RL effective. Based on these findings, we propose a boundary-aware recipe and instantiate it by RL post-training an OctoMed-initialized model on a small, balanced subset of PMC multiple-choice VQA, achieving strong average performance across six medical VQA benchmarks.</p> - https://arxiv.org/abs/2603.01301 - Sun, 01 Mar 2026 22:16:19 +0000 + NLE: Non-autoregressive LLM-based ASR by Transcript Editing + https://arxiv.org/abs/2603.08397 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08397.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Avihu Dekel, Samuel Thomas, Takashi Fukada, George Saon</p><p><b>Upvotes:</b> 19</p><p><b>Summary:</b> While autoregressive (AR) LLM-based ASR systems achieve strong accuracy, their sequential decoding limits parallelism and incurs high latency. We propose NLE, a non-autoregressive (NAR) approach that formulates speech recognition as conditional transcript editing, enabling fully parallel prediction. NLE extracts acoustic embeddings and an initial hypothesis from a pretrained speech encoder, then refines the hypothesis using a bidirectional LLM editor trained with a latent alignment objective. An interleaved padding strategy exploits the identity mapping bias of Transformers, allowing the model to focus on corrections rather than full reconstruction. On the Open ASR leaderboard, NLE++ achieves 5.67% average WER with an RTFx (inverse real-time factor) of 1630. In single-utterance scenarios, NLE achieves 27x speedup over the AR baseline, making it suitable for real-time applications.</p> + https://arxiv.org/abs/2603.08397 + Mon, 09 Mar 2026 13:56:52 +0000 - SciDER: Scientific Data-centric End-to-end Researcher - https://arxiv.org/abs/2603.01421 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01421.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ke Lin, Yilin Lu, Shreyas Bhat, Xuehang Guo, Junier Oliva, Qingyun Wang</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Automated scientific discovery with large language models is transforming the research lifecycle from ideation to experimentation, yet existing agents struggle to autonomously process raw data collected from scientific experiments. We introduce SciDER, a data-centric end-to-end system that automates the research lifecycle. Unlike traditional frameworks, our specialized agents collaboratively parse and analyze raw scientific data, generate hypotheses and experimental designs grounded in specific data characteristics, and write and execute corresponding code. Evaluation on three benchmarks shows SciDER excels in specialized data-driven scientific discovery and outperforms general-purpose agents and state-of-the-art models through its self-evolving memory and critic-led feedback loop. Distributed as a modular Python package, we also provide easy-to-use PyPI packages with a lightweight web interface to accelerate autonomous, data-driven research and aim to be accessible to all researchers and developers.</p> - https://arxiv.org/abs/2603.01421 - Mon, 02 Mar 2026 03:53:20 +0000 + RetroAgent: From Solving to Evolving via Retrospective Dual Intrinsic Feedback + https://arxiv.org/abs/2603.08561 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08561.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiaoying Zhang, Zichen Liu, Yipeng Zhang, Xia Hu, Wenqi Shao</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Large language model (LLM)-based agents trained with reinforcement learning (RL) have shown strong potential on complex interactive tasks. However, standard RL paradigms favor static problem-solving over continuous adaptation: agents often converge to suboptimal strategies due to insufficient exploration, while learned knowledge remains implicit within parameters rather than explicitly retrievable, limiting effective experiential learning. To address these limitations, we introduce RetroAgent, an online RL framework that empowers agents to master complex interactive environments not just by solving, but by evolving. Concretely, RetroAgent features a hindsight self-reflection mechanism that produces dual intrinsic feedback: (1) intrinsic numerical feedback that that tracks incremental subtask completion relative to prior attempts, rewarding promising explorations, and (2) intrinsic language feedback that distills reusable lessons into a memory buffer, retrieved via our proposed Similarity & Utility-Aware Upper Confidence Bound (SimUtil-UCB) strategy balancing relevance, utility, and exploration to effectively leverage past experiences. Extensive experiments on two model families across four challenging agentic tasks demonstrate that RetroAgent significantly outperforms existing methods, achieving state-of-the-art results -- e.g., surpassing Group Relative Policy Optimization (GRPO)-trained agents by +18.3% on ALFWorld, +15.4% on WebShop, +27.1% on Sokoban, and +8.9% on MineSweeper -- while exhibiting strong test-time adaptation and generalization to out-of-distribution scenarios.</p> + https://arxiv.org/abs/2603.08561 + Mon, 09 Mar 2026 16:23:33 +0000 - LaSER: Internalizing Explicit Reasoning into Latent Space for Dense Retrieval - https://arxiv.org/abs/2603.01425 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01425.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiajie Jin, Yanzhao Zhang, Mingxin Li, Dingkun Long, Pengjun Xie, Yutao Zhu, Zhicheng Dou</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> LLMs have fundamentally transformed dense retrieval, upgrading backbones from discriminative encoders to generative architectures. However, a critical disconnect remains: while LLMs possess strong reasoning capabilities, current retrievers predominantly utilize them as static encoders, leaving their potential for complex reasoning unexplored. To address this, existing approaches typically adopt rewrite-then-retrieve pipelines to generate explicit CoT rationales before retrieval. However, this incurs prohibitive latency. In this paper, we propose LaSER, a novel self-distillation framework that internalizes explicit reasoning into the latent space of dense retrievers. Operating on a shared LLM backbone, LaSER introduces a dual-view training mechanism: an Explicit view that explicitly encodes ground-truth reasoning paths, and a Latent view that performs implicit latent thinking. To bridge the gap between these views, we design a multi-grained alignment strategy. Beyond standard output alignment, we introduce a trajectory alignment mechanism that synchronizes the intermediate latent states of the latent path with the semantic progression of the explicit reasoning segments. This allows the retriever to think silently and effectively without autoregressive text generation. Extensive experiments on both in-domain and out-of-domain reasoning-intensive benchmarks demonstrate that LaSER significantly outperforms state-of-the-art baselines. Furthermore, analyses across diverse backbones and model scales validate the robustness of our approach, confirming that our unified learning framework is essential for eliciting effective latent thinking. Our method successfully combines the reasoning depth of explicit CoT pipelines with the inference efficiency of standard dense retrievers.</p> - https://arxiv.org/abs/2603.01425 - Mon, 02 Mar 2026 04:11:18 +0000 + CARE-Edit: Condition-Aware Routing of Experts for Contextual Image Editing + https://arxiv.org/abs/2603.08589 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08589.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yucheng Wang, Zedong Wang, Yuetong Wu, Yue Ma, Dan Xu</p><p><b>Upvotes:</b> 34</p><p><b>Summary:</b> Unified diffusion editors often rely on a fixed, shared backbone for diverse tasks, suffering from task interference and poor adaptation to heterogeneous demands (e.g., local vs global, semantic vs photometric). In particular, prevalent ControlNet and OmniControl variants combine multiple conditioning signals (e.g., text, mask, reference) via static concatenation or additive adapters which cannot dynamically prioritize or suppress conflicting modalities, thus resulting in artifacts like color bleeding across mask boundaries, identity or style drift, and unpredictable behavior under multi-condition inputs. To address this, we propose Condition-Aware Routing of Experts (CARE-Edit) that aligns model computation with specific editing competencies. At its core, a lightweight latent-attention router assigns encoded diffusion tokens to four specialized experts--Text, Mask, Reference, and Base--based on multi-modal conditions and diffusion timesteps: (i) a Mask Repaint module first refines coarse user-defined masks for precise spatial guidance; (ii) the router applies sparse top-K selection to dynamically allocate computation to the most relevant experts; (iii) a Latent Mixture module subsequently fuses expert outputs, coherently integrating semantic, spatial, and stylistic information to the base images. Experiments validate CARE-Edit's strong performance on contextual editing tasks, including erasure, replacement, text-driven edits, and style transfer. Empirical analysis further reveals task-specific behavior of specialized experts, showcasing the importance of dynamic, condition-aware processing to mitigate multi-condition conflicts.</p> + https://arxiv.org/abs/2603.08589 + Mon, 09 Mar 2026 16:40:47 +0000 - PhotoBench: Beyond Visual Matching Towards Personalized Intent-Driven Photo Retrieval - https://arxiv.org/abs/2603.01493 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01493.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tianyi Xu, Rong Shan, Junjie Wu, Jiadeng Huang, Teng Wang, Jiachen Zhu, Wenteng Chen, Minxin Tu, Quantao Dou, Zhaoxiang Wang, Changwang Zhang, Weinan Zhang, Jun Wang, Jianghao Lin</p><p><b>Upvotes:</b> 20</p><p><b>Summary:</b> Personal photo albums are not merely collections of static images but living, ecological archives defined by temporal continuity, social entanglement, and rich metadata, which makes the personalized photo retrieval non-trivial. However, existing retrieval benchmarks rely heavily on context-isolated web snapshots, failing to capture the multi-source reasoning required to resolve authentic, intent-driven user queries. To bridge this gap, we introduce PhotoBench, the first benchmark constructed from authentic, personal albums. It is designed to shift the paradigm from visual matching to personalized multi-source intent-driven reasoning. Based on a rigorous multi-source profiling framework, which integrates visual semantics, spatial-temporal metadata, social identity, and temporal events for each image, we synthesize complex intent-driven queries rooted in users' life trajectories. Extensive evaluation on PhotoBench exposes two critical limitations: the modality gap, where unified embedding models collapse on non-visual constraints, and the source fusion paradox, where agentic systems perform poor tool orchestration. These findings indicate that the next frontier in personal multimodal retrieval lies beyond unified embeddings, necessitating robust agentic reasoning systems capable of precise constraint satisfaction and multi-source fusion. Our PhotoBench is available.</p> - https://arxiv.org/abs/2603.01493 - Mon, 02 Mar 2026 06:02:40 +0000 + CAST: Modeling Visual State Transitions for Consistent Video Retrieval + https://arxiv.org/abs/2603.08648 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08648.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yanqing Liu, Yingcheng Liu, Fanghong Dong, Budianto Budianto, Cihang Xie, Yan Jiao</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> As video content creation shifts toward long-form narratives, composing short clips into coherent storylines becomes increasingly important. However, prevailing retrieval formulations remain context-agnostic at inference time, prioritizing local semantic alignment while neglecting state and identity consistency. To address this structural limitation, we formalize the task of Consistent Video Retrieval (CVR) and introduce a diagnostic benchmark spanning YouCook2, COIN, and CrossTask. We propose CAST (Context-Aware State Transition), a lightweight, plug-and-play adapter compatible with diverse frozen vision-language embedding spaces. By predicting a state-conditioned residual update (Δ) from visual history, CAST introduces an explicit inductive bias for latent state evolution. Extensive experiments show that CAST improves performance on YouCook2 and CrossTask, remains competitive on COIN, and consistently outperforms zero-shot baselines across diverse foundation backbones. Furthermore, CAST provides a useful reranking signal for black-box video generation candidates (e.g., from Veo), promoting more temporally coherent continuations.</p> + https://arxiv.org/abs/2603.08648 + Mon, 09 Mar 2026 17:26:26 +0000 - RubricBench: Aligning Model-Generated Rubrics with Human Standards - https://arxiv.org/abs/2603.01562 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01562.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qiyuan Zhang, Junyi Zhou, Yufei Wang, Fuyuan Lyu, Yidong Ming, Can Xu, Qingfeng Sun, Kai Zheng, Peng Kang, Xue Liu, Chen Ma</p><p><b>Upvotes:</b> 51</p><p><b>Summary:</b> As Large Language Model (LLM) alignment evolves from simple completions to complex, highly sophisticated generation, Reward Models are increasingly shifting toward rubric-guided evaluation to mitigate surface-level biases. However, the community lacks a unified benchmark to assess this evaluation paradigm, as existing benchmarks lack both the discriminative complexity and the ground-truth rubric annotations required for rigorous analysis. To bridge this gap, we introduce RubricBench, a curated benchmark with 1,147 pairwise comparisons specifically designed to assess the reliability of rubric-based evaluation. Our construction employs a multi-dimensional filtration pipeline to target hard samples featuring nuanced input complexity and misleading surface bias, augmenting each with expert-annotated, atomic rubrics derived strictly from instructions. Comprehensive experiments reveal a substantial capability gap between human-annotated and model-generated rubrics, indicating that even state-of-the-art models struggle to autonomously specify valid evaluation criteria, lagging considerably behind human-guided performance.</p> - https://arxiv.org/abs/2603.01562 - Mon, 02 Mar 2026 07:39:49 +0000 + CoCo: Code as CoT for Text-to-Image Preview and Rare Concept Generation + https://arxiv.org/abs/2603.08652 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08652.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haodong Li, Chunmei Qing, Huanyu Zhang, Dongzhi Jiang, Yihang Zou, Hongbo Peng, Dingming Li, Yuhong Dai, ZePeng Lin, Juanxi Tian, Yi Zhou, Siqi Dai, Jingwei Wu</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> Recent advancements in Unified Multimodal Models (UMMs) have significantly advanced text-to-image (T2I) generation, particularly through the integration of Chain-of-Thought (CoT) reasoning. However, existing CoT-based T2I methods largely rely on abstract natural-language planning, which lacks the precision required for complex spatial layouts, structured visual elements, and dense textual content. In this work, we propose CoCo (Code-as-CoT), a code-driven reasoning framework that represents the reasoning process as executable code, enabling explicit and verifiable intermediate planning for image generation. Given a text prompt, CoCo first generates executable code that specifies the structural layout of the scene, which is then executed in a sandboxed environment to render a deterministic draft image. The model subsequently refines this draft through fine-grained image editing to produce the final high-fidelity result. To support this training paradigm, we construct CoCo-10K, a curated dataset containing structured draft-final image pairs designed to teach both structured draft construction and corrective visual refinement. Empirical evaluations on StructT2IBench, OneIG-Bench, and LongText-Bench show that CoCo achieves improvements of +68.83%, +54.8%, and +41.23% over direct generation, while also outperforming other generation methods empowered by CoT. These results demonstrate that executable code is an effective and reliable reasoning paradigm for precise, controllable, and structured text-to-image generation. The code is available at: https://github.com/micky-li-hd/CoCo</p> + https://arxiv.org/abs/2603.08652 + Mon, 09 Mar 2026 17:31:16 +0000 - Beyond Length Scaling: Synergizing Breadth and Depth for Generative Reward Models - https://arxiv.org/abs/2603.01571 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01571.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qiyuan Zhang, Yufei Wang, Tianhe Wu, Can Xu, Qingfeng Sun, Kai Zheng, Xue Liu, Chen Ma</p><p><b>Upvotes:</b> 32</p><p><b>Summary:</b> Recent advancements in Generative Reward Models (GRMs) have demonstrated that scaling the length of Chain-of-Thought (CoT) reasoning considerably enhances the reliability of evaluation. However, current works predominantly rely on unstructured length scaling, ignoring the divergent efficacy of different reasoning mechanisms: Breadth-CoT (B-CoT, i.e., multi-dimensional principle coverage) and Depth-CoT (D-CoT, i.e., substantive judgment soundness). To address this, we introduce Mix-GRM, a framework that reconfigures raw rationales into structured B-CoT and D-CoT through a modular synthesis pipeline, subsequently employing Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR) to internalize and optimize these mechanisms. Comprehensive experiments demonstrate that Mix-GRM establishes a new state-of-the-art across five benchmarks, surpassing leading open-source RMs by an average of 8.2\%. Our results reveal a clear divergence in reasoning: B-CoT benefits subjective preference tasks, whereas D-CoT excels in objective correctness tasks. Consequently, misaligning the reasoning mechanism with the task directly degrades performance. Furthermore, we demonstrate that RLVR acts as a switching amplifier, inducing an emergent polarization where the model spontaneously allocates its reasoning style to match task demands. The synthesized data and models are released at https://huggingface.co/collections/DonJoey/mix-grm{Hugging Face}, and the code is released at https://github.com/Don-Joey/Mix-GRM{Github}.</p> - https://arxiv.org/abs/2603.01571 - Mon, 02 Mar 2026 07:54:29 +0000 + How Far Can Unsupervised RLVR Scale LLM Training? + https://arxiv.org/abs/2603.08660 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08660.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Bingxiang He, Yuxin Zuo, Zeyuan Liu, Shangziqi Zhao, Zixuan Fu, Junlin Yang, Cheng Qian, Kaiyan Zhang, Yuchen Fan, Ganqu Cui, Xiusi Chen, Youbang Sun, Xingtai Lv, Xuekai Zhu, Li Sheng, Ran Li, Huan-ang Gao, Yuchen Zhang, Bowen Zhou, Zhiyuan Liu, Ning Ding</p><p><b>Upvotes:</b> 47</p><p><b>Summary:</b> Unsupervised reinforcement learning with verifiable rewards (URLVR) offers a pathway to scale LLM training beyond the supervision bottleneck by deriving rewards without ground truth labels. Recent works leverage model intrinsic signals, showing promising early gains, yet their potential and limitations remain unclear. In this work, we revisit URLVR and provide a comprehensive analysis spanning taxonomy, theory and extensive experiments. We first classify URLVR methods into intrinsic versus external based on reward sources, then establish a unified theoretical framework revealing that all intrinsic methods converge toward sharpening the model's initial distribution This sharpening mechanism succeeds when initial confidence aligns with correctness but fails catastrophically when misaligned. Through systematic experiments, we show intrinsic rewards consistently follow a rise-then-fall pattern across methods, with collapse timing determined by model prior rather than engineering choices. Despite these scaling limits, we find intrinsic rewards remain valuable in test-time training on small datasets, and propose Model Collapse Step to measure model prior, serving as a practical indicator for RL trainability. Finally, we explore external reward methods that ground verification in computational asymmetries, showing preliminary evidence they may escape the confidence-correctness ceiling. Our findings chart boundaries for intrinsic URLVR while motivating paths toward scalable alternatives.</p> + https://arxiv.org/abs/2603.08660 + Mon, 09 Mar 2026 17:38:11 +0000 - Surgical Post-Training: Cutting Errors, Keeping Knowledge - https://arxiv.org/abs/2603.01683 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01683.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wenye Lin, Kai Han</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Enhancing the reasoning capabilities of Large Language Models (LLMs) via post-training is often constrained by the trade-off between efficiency and catastrophic forgetting. While prior research emphasizes the role of on-policy data in mitigating forgetting, we uncover--and validate both theoretically and empirically--an overlooked yet critical mechanism: the implicit regularization inherent in Direct Preference Optimization's (DPO) reward estimate. This motivates our Surgical Post-Training (SPoT), a new paradigm designed to optimize reasoning efficiently while preserving learned prior knowledge. SPoT consists of: (1) a data rectification pipeline that employs an Oracle to surgically correct erroneous steps via minimal edits, generating data proximal to the model's distribution; and (2) a reward-based binary cross-entropy objective. Unlike the relative ranking in DPO, this objective treats reasoning correctness as a binary classification problem, enforcing decoupled supervision signals. Empirically, with only 4k rectified math data pairs, SPoT improves Qwen3-8B's accuracy by 6.2% on average across in-domain and OOD tasks, requiring merely 28 minutes of training on 8x H800 GPUs. Code: https://github.com/Visual-AI/SPoT</p> - https://arxiv.org/abs/2603.01683 - Mon, 02 Mar 2026 10:12:56 +0000 + HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising + https://arxiv.org/abs/2603.08703 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08703.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kai Zou, Dian Zheng, Hongbo Liu, Tiankai Hang, Bin Liu, Nenghai Yu</p><p><b>Upvotes:</b> 29</p><p><b>Summary:</b> Autoregressive (AR) diffusion offers a promising framework for generating videos of theoretically infinite length. However, a major challenge is maintaining temporal continuity while preventing the progressive quality degradation caused by error accumulation. To ensure continuity, existing methods typically condition on highly denoised contexts; yet, this practice propagates prediction errors with high certainty, thereby exacerbating degradation. In this paper, we argue that a highly clean context is unnecessary. Drawing inspiration from bidirectional diffusion models, which denoise frames at a shared noise level while maintaining coherence, we propose that conditioning on context at the same noise level as the current block provides sufficient signal for temporal consistency while effectively mitigating error propagation. Building on this insight, we propose HiAR, a hierarchical denoising framework that reverses the conventional generation order: instead of completing each block sequentially, it performs causal generation across all blocks at every denoising step, so that each block is always conditioned on context at the same noise level. This hierarchy naturally admits pipelined parallel inference, yielding a 1.8 wall-clock speedup in our 4-step setting. We further observe that self-rollout distillation under this paradigm amplifies a low-motion shortcut inherent to the mode-seeking reverse-KL objective. To counteract this, we introduce a forward-KL regulariser in bidirectional-attention mode, which preserves motion diversity for causal inference without interfering with the distillation loss. On VBench (20s generation), HiAR achieves the best overall score and the lowest temporal drift among all compared methods.</p> + https://arxiv.org/abs/2603.08703 + Mon, 09 Mar 2026 17:58:16 +0000 - Legal RAG Bench: an end-to-end benchmark for legal RAG - https://arxiv.org/abs/2603.01710 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01710.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Abdur-Rahman Butler, Umar Butler</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> We introduce Legal RAG Bench, a benchmark and evaluation methodology for assessing the end-to-end performance of legal RAG systems. As a benchmark, Legal RAG Bench consists of 4,876 passages from the Victorian Criminal Charge Book alongside 100 complex, hand-crafted questions demanding expert knowledge of criminal law and procedure. Both long-form answers and supporting passages are provided. As an evaluation methodology, Legal RAG Bench leverages a full factorial design and novel hierarchical error decomposition framework, enabling apples-to-apples comparisons of the contributions of retrieval and reasoning models in RAG. We evaluate three state-of-the-art embedding models (Isaacus' Kanon 2 Embedder, Google's Gemini Embedding 001, and OpenAI's Text Embedding 3 Large) and two frontier LLMs (Gemini 3.1 Pro and GPT-5.2), finding that information retrieval is the primary driver of legal RAG performance, with LLMs exerting a more moderate effect on correctness and groundedness. Kanon 2 Embedder, in particular, had the largest positive impact on performance, improving average correctness by 17.5 points, groundedness by 4.5 points, and retrieval accuracy by 34 points. We observe that many errors attributed to hallucinations in legal RAG systems are in fact triggered by retrieval failures, concluding that retrieval sets the ceiling for the performance of many modern legal RAG systems. We document why and how we built Legal RAG Bench alongside the results of our evaluations. We also openly release our code and data to assist with reproduction of our findings.</p> - https://arxiv.org/abs/2603.01710 - Mon, 02 Mar 2026 10:34:28 +0000 + Agentic Critical Training + https://arxiv.org/abs/2603.08706 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08706.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Weize Liu, Minghui Liu, Sy-Tuyen Ho, Souradip Chakraborty, Xiyao Wang, Furong Huang</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Training large language models (LLMs) as autonomous agents often begins with imitation learning, but it only teaches agents what to do without understanding why: agents never contrast successful actions against suboptimal alternatives and thus lack awareness of action quality. Recent approaches attempt to address this by introducing self-reflection supervision derived from contrasts between expert and alternative actions. However, the training paradigm fundamentally remains imitation learning: the model imitates pre-constructed reflection text rather than learning to reason autonomously. We propose Agentic Critical Training (ACT), a reinforcement learning paradigm that trains agents to identify the better action among alternatives. By rewarding whether the model's judgment is correct, ACT drives the model to autonomously develop reasoning about action quality, producing genuine self-reflection rather than imitating it. Across three challenging agent benchmarks, ACT consistently improves agent performance when combined with different post-training methods. It achieves an average improvement of 5.07 points over imitation learning and 4.62 points over reinforcement learning. Compared to approaches that inject reflection capability through knowledge distillation, ACT also demonstrates clear advantages, yielding an average improvement of 2.42 points. Moreover, ACT enables strong out-of-distribution generalization on agentic benchmarks and improves performance on general reasoning benchmarks without any reasoning-specific training data, highlighting the value of our method. These results suggest that ACT is a promising path toward developing more reflective and capable LLM agents.</p> + https://arxiv.org/abs/2603.08706 + Mon, 09 Mar 2026 17:58:56 +0000 - OpenAutoNLU: Open Source AutoML Library for NLU - https://arxiv.org/abs/2603.01824 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01824.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Grigory Arshinov, Aleksandr Boriskin, Sergey Senichev, Ayaz Zaripov, Daria Galimzianova, Daniil Karpov, Leonid Sanochkin</p><p><b>Upvotes:</b> 44</p><p><b>Summary:</b> OpenAutoNLU is an open-source automated machine learning library for natural language understanding (NLU) tasks, covering both text classification and named entity recognition (NER). Unlike existing solutions, we introduce data-aware training regime selection that requires no manual configuration from the user. The library also provides integrated data quality diagnostics, configurable out-of-distribution (OOD) detection, and large language model (LLM) features, all within a minimal lowcode API. The demo app is accessible here https://openautonlu.dev.</p> - https://arxiv.org/abs/2603.01824 - Mon, 02 Mar 2026 12:56:54 +0000 + Scale Space Diffusion + https://arxiv.org/abs/2603.08709 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08709.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Soumik Mukhopadhyay, Prateksha Udhayanan, Abhinav Shrivastava</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Diffusion models degrade images through noise, and reversing this process reveals an information hierarchy across timesteps. Scale-space theory exhibits a similar hierarchy via low-pass filtering. We formalize this connection and show that highly noisy diffusion states contain no more information than small, downsampled images - raising the question of why they must be processed at full resolution. To address this, we fuse scale spaces into the diffusion process by formulating a family of diffusion models with generalized linear degradations and practical implementations. Using downsampling as the degradation yields our proposed Scale Space Diffusion. To support Scale Space Diffusion, we introduce Flexi-UNet, a UNet variant that performs resolution-preserving and resolution-increasing denoising using only the necessary parts of the network. We evaluate our framework on CelebA and ImageNet and analyze its scaling behavior across resolutions and network depths. Our project website ( https://prateksha.github.io/projects/scale-space-diffusion/ ) is available publicly.</p> + https://arxiv.org/abs/2603.08709 + Mon, 09 Mar 2026 17:59:42 +0000 - FireRed-OCR Technical Report - https://arxiv.org/abs/2603.01840 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01840.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hao Wu, Haoran Lou, Xinyue Li, Zuodong Zhong, Zhaojun Sun, Phellon Chen, Xuanhe Zhou, Kai Zuo, Yibo Chen, Xu Tang, Yao Hu, Boxiang Zhou, Jian Wu, Yongji Wu, Wenxin Yu, Yingmiao Liu, Yuhao Huang, Manjie Xu, Gang Liu, Yidong Ma, Zhichao Sun, Changhao Qiao</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> We present FireRed-OCR, a systematic framework to specialize general VLMs into high-performance OCR models. Large Vision-Language Models (VLMs) have demonstrated impressive general capabilities but frequently suffer from ``structural hallucination'' when processing complex documents, limiting their utility in industrial OCR applications. In this paper, we introduce FireRed-OCR, a novel framework designed to transform general-purpose VLMs (based on Qwen3-VL) into pixel-precise structural document parsing experts. To address the scarcity of high-quality structured data, we construct a ``Geometry + Semantics'' Data Factory. Unlike traditional random sampling, our pipeline leverages geometric feature clustering and multi-dimensional tagging to synthesize and curate a highly balanced dataset, effectively handling long-tail layouts and rare document types. Furthermore, we propose a Three-Stage Progressive Training strategy that guides the model from pixel-level perception to logical structure generation. This curriculum includes: (1) Multi-task Pre-alignment to ground the model's understanding of document structure; (2) Specialized SFT for standardizing full-image Markdown output; and (3) Format-Constrained Group Relative Policy Optimization (GRPO), which utilizes reinforcement learning to enforce strict syntactic validity and structural integrity (e.g., table closure, formula syntax). Extensive evaluations on OmniDocBench v1.5 demonstrate that FireRed-OCR achieves state-of-the-art performance with an overall score of 92.94\%, significantly outperforming strong baselines such as DeepSeek-OCR 2 and OCRVerse across text, formula, table, and reading order metrics. We open-source our code and model weights to facilitate the ``General VLM to Specialized Structural Expert'' paradigm.</p> - https://arxiv.org/abs/2603.01840 - Mon, 02 Mar 2026 13:19:23 +0000 + Test-Driven AI Agent Definition (TDAD): Compiling Tool-Using Agents from Behavioral Specifications + https://arxiv.org/abs/2603.08806 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08806.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tzafrir Rehan</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> We present Test-Driven AI Agent Definition (TDAD), a methodology that treats agent prompts as compiled artifacts: engineers provide behavioral specifications, a coding agent converts them into executable tests, and a second coding agent iteratively refines the prompt until tests pass. Deploying tool-using LLM agents in production requires measurable behavioral compliance that current development practices cannot provide. Small prompt changes cause silent regressions, tool misuse goes undetected, and policy violations emerge only after deployment. To mitigate specification gaming, TDAD introduces three mechanisms: (1) visible/hidden test splits that withhold evaluation tests during compilation, (2) semantic mutation testing via a post-compilation agent that generates plausible faulty prompt variants, with the harness measuring whether the test suite detects them, and (3) spec evolution scenarios that quantify regression safety when requirements change. We evaluate TDAD on SpecSuite-Core, a benchmark of four deeply-specified agents spanning policy compliance, grounded analytics, runbook adherence, and deterministic enforcement. Across 24 independent trials, TDAD achieves 92% v1 compilation success with 97% mean hidden pass rate; evolved specifications compile at 58%, with most failed runs passing all visible tests except 1-2, and show 86-100% mutation scores, 78% v2 hidden pass rate, and 97% regression safety scores. The implementation is available as an open benchmark at https://github.com/f-labs-io/tdad-paper-code.</p> + https://arxiv.org/abs/2603.08806 + Mon, 09 Mar 2026 18:04:54 +0000 - Agentic Code Reasoning - https://arxiv.org/abs/2603.01896 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01896.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shubham Ugare, Satish Chandra</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Can LLM agents explore codebases and reason about code semantics without executing the code? We study this capability, which we call agentic code reasoning, and introduce semi-formal reasoning: a structured prompting methodology that requires agents to construct explicit premises, trace execution paths, and derive formal conclusions. Unlike unstructured chain-of-thought, semi-formal reasoning acts as a certificate: the agent cannot skip cases or make unsupported claims. We evaluate across three tasks (patch equivalence verification, fault localization, and code question answering) and show that semi-formal reasoning consistently improves accuracy on all of them. For patch equivalence, accuracy improves from 78% to 88% on curated examples and reaches 93% on real-world agent-generated patches, approaching the reliability needed for execution-free RL reward signals. For code question answering on RubberDuckBench Mohammad et al. (2026), semi-formal reasoning achieves 87% accuracy. For fault localization on Defects4J Just et al. (2014), semi-formal reasoning improves Top-5 accuracy by 5 percentage points over standard reasoning. These results demonstrate that structured agentic reasoning enables meaningful semantic code analysis without execution, opening practical applications in RL training pipelines, code review, and static program analysis.</p> - https://arxiv.org/abs/2603.01896 - Mon, 02 Mar 2026 14:17:06 +0000 + Training Language Models via Neural Cellular Automata + https://arxiv.org/abs/2603.10055 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10055.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dan Lee, Seungwook Han, Akarsh Kumar, Pulkit Agrawal</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Pre-training is crucial for large language models (LLMs), as it is when most representations and capabilities are acquired. However, natural language pre-training has problems: high-quality text is finite, it contains human biases, and it entangles knowledge with reasoning. This raises a fundamental question: is natural language the only path to intelligence? We propose using neural cellular automata (NCA) to generate synthetic, non-linguistic data for pre-pre-training LLMs--training on synthetic-then-natural language. NCA data exhibits rich spatiotemporal structure and statistics resembling natural language while being controllable and cheap to generate at scale. We find that pre-pre-training on only 164M NCA tokens improves downstream language modeling by up to 6% and accelerates convergence by up to 1.6x. Surprisingly, this even outperforms pre-pre-training on 1.6B tokens of natural language from Common Crawl with more compute. These gains also transfer to reasoning benchmarks, including GSM8K, HumanEval, and BigBench-Lite. Investigating what drives transfer, we find that attention layers are the most transferable, and that optimal NCA complexity varies by domain: code benefits from simpler dynamics, while math and web text favor more complex ones. These results enable systematic tuning of the synthetic distribution to target domains. More broadly, our work opens a path toward more efficient models with fully synthetic pre-training.</p> + https://arxiv.org/abs/2603.10055 + Mon, 09 Mar 2026 18:14:26 +0000 - Efficient RLVR Training via Weighted Mutual Information Data Selection - https://arxiv.org/abs/2603.01907 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01907.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xinyu Zhou, Boyu Zhu, Haotian Zhang, Huiming Wang, Zhijiang Guo</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Reinforcement learning (RL) plays a central role in improving the reasoning and alignment of large language models, yet its efficiency critically depends on how training data are selected. Existing online selection strategies predominantly rely on difficulty-based heuristics, favouring datapoints with intermediate success rates, implicitly equating difficulty with informativeness and neglecting epistemic uncertainty arising from limited evidence. We introduce InSight, an INformation-guided data SamplInG metHod for RL Training, grounded in a weighted mutual information objective. By modeling data outcomes with Bayesian latent success rates, we show that expected uncertainty reduction decomposes into complementary difficulty- and evidence-dependent components, revealing a fundamental limitation of difficulty-only selection. Leveraging this observation, InSight constructs a stable acquisition score based on the mean belief of datapoints' success rather than noisy sampled outcomes, and naturally extends to multi-rollout settings common in reinforcement learning with verifiable rewards (RLVR). Extensive experiments demonstrate that InSight consistently achieves state-of-the-art performance and improves training efficiency, including a +1.41 average gain on Planning & Mathmatics benchmarks, +1.01 improvement on general reasoning, and up to ~2.2x acceleration, with negligible additional computational overhead.</p> - https://arxiv.org/abs/2603.01907 - Mon, 02 Mar 2026 14:25:07 +0000 + Fish Audio S2 Technical Report + https://arxiv.org/abs/2603.08823 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.08823.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shijia Liao, Yuxuan Wang, Songting Liu, Yifan Cheng, Ruoyi Zhang, Tianyu Li, Shidong Li, Yisheng Zheng, Xingwei Liu, Qingzheng Wang, Zhizhuo Zhou, Jiahua Liu, Xin Chen, Dawei Han</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> We introduce Fish Audio S2, an open-sourced text-to-speech system featuring multi-speaker, multi-turn generation, and, most importantly, instruction-following control via natural-language descriptions. To scale training, we develop a multi-stage training recipe together with a staged data pipeline covering video captioning and speech captioning, voice-quality assessment, and reward modeling. To push the frontier of open-source TTS, we release our model weights, fine-tuning code, and an SGLang-based inference engine. The inference engine is production-ready for streaming, achieving an RTF of 0.195 and a time-to-first-audio below 100 ms.Our code and weights are available on GitHub (https://github.com/fishaudio/fish-speech) and Hugging Face (https://huggingface.co/fishaudio/s2-pro). We highly encourage readers to visit https://fish.audio to try custom voices.</p> + https://arxiv.org/abs/2603.08823 + Mon, 09 Mar 2026 18:34:33 +0000 - CoVe: Training Interactive Tool-Use Agents via Constraint-Guided Verification - https://arxiv.org/abs/2603.01940 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01940.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jinpeng Chen, Cheng Gong, Hanbo Li, Ziru Liu, Zichen Tian, Xinyu Fu, Shi Wu, Chenyang Zhang, Wu Zhang, Suiyun Zhang, Dandan Tu, Rui Liu</p><p><b>Upvotes:</b> 21</p><p><b>Summary:</b> Developing multi-turn interactive tool-use agents is challenging because real-world user needs are often complex and ambiguous, yet agents must execute deterministic actions to satisfy them. To address this gap, we introduce CoVe (Constraint-Verification), a post-training data synthesis framework designed for training interactive tool-use agents while ensuring both data complexity and correctness. CoVe begins by defining explicit task constraints, which serve a dual role: they guide the generation of complex trajectories and act as deterministic verifiers for assessing trajectory quality. This enables the creation of high-quality training trajectories for supervised fine-tuning (SFT) and the derivation of accurate reward signals for reinforcement learning (RL). Our evaluation on the challenging τ^2-bench benchmark demonstrates the effectiveness of the framework. Notably, our compact CoVe-4B model achieves success rates of 43.0\% and 59.4\% in the Airline and Retail domains, respectively; its overall performance significantly outperforms strong baselines of similar scale and remains competitive with models up to 17times its size. These results indicate that CoVe provides an effective and efficient pathway for synthesizing training data for state-of-the-art interactive tool-use agents. To support future research, we open-source our code, trained model, and the full set of 12K high-quality trajectories used for training.</p> - https://arxiv.org/abs/2603.01940 - Mon, 02 Mar 2026 14:56:35 +0000 + Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs + https://arxiv.org/abs/2603.09095 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09095.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kaiser Sun, Xiaochuang Yuan, Hongjun Liu, Chen Zhao, Cheng Zhang, Mark Dredze, Fan Bai</p><p><b>Upvotes:</b> 25</p><p><b>Summary:</b> Multimodal large language models (MLLMs) can process text presented as images, yet they often perform worse than when the same content is provided as textual tokens. We systematically diagnose this "modality gap" by evaluating seven MLLMs across seven benchmarks in five input modes, spanning both synthetically rendered text and realistic document images from arXiv PDFs to Wikipedia pages. We find that the modality gap is task- and data-dependent. For example, math tasks degrade by over 60 points on synthetic renderings, while natural document images often match or exceed text-mode performance. Rendering choices such as font and resolution are strong confounds, with font alone swinging accuracy by up to 47 percentage points. To understand this, we conduct a grounded-theory error analysis of over 4,000 examples, revealing that image mode selectively amplifies reading errors (calculation and formatting failures) while leaving knowledge and reasoning errors largely unchanged, and that some models exhibit a chain-of-thought reasoning collapse under visual input. Motivated by these findings, we propose a self-distillation method that trains the model on its own pure text reasoning traces paired with image inputs, raising image-mode accuracy on GSM8K from 30.71% to 92.72% and transferring to unseen benchmarks without catastrophic forgetting. Overall, our study provides a systematic understanding of the modality gap and suggests a practical path toward improving visual text understanding in multimodal language models.</p> + https://arxiv.org/abs/2603.09095 + Tue, 10 Mar 2026 02:14:23 +0000 - CharacterFlywheel: Scaling Iterative Improvement of Engaging and Steerable LLMs in Production - https://arxiv.org/abs/2603.01973 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.01973.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yixin Nie, Lin Guan, Zhongyao Ma, Anchit Gupta, Yipin Zhou, Xiao Li, Zhengping Zhou, Raymond Zeng, Gelin Zhou, Shigan Chu, Ajay Thampi, Wancen Mu, Nathan Shuster, Ketong Wang, Lin Chen, Jason Brewer, Derek Hao Hu, Alexander McCauley, Jason Weston, Sem Park, Na Zhang, Kevin Tang</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> This report presents CharacterFlywheel, an iterative flywheel process for improving large language models (LLMs) in production social chat applications across Instagram, WhatsApp, and Messenger. Starting from LLaMA 3.1, we refined models across 15 generations using data from both internal and external real-user traffic. Through continuous deployments from July 2024 to April 2025, we conducted controlled 7-day A/B tests showing consistent engagement improvements: 7 of 8 newly deployed models demonstrated positive lift over the baseline, with the strongest performers achieving up to 8.8% improvement in engagement breadth and 19.4% in engagement depth. We also observed substantial gains in steerability, with instruction following increasing from 59.2% to 84.8% and instruction violations decreasing from 26.6% to 5.8%. We detail the CharacterFlywheel process which integrates data curation, reward modeling to estimate and interpolate the landscape of engagement metrics, supervised fine-tuning (SFT), reinforcement learning (RL), and both offline and online evaluation to ensure reliable progress at each optimization step. We also discuss our methods for overfitting prevention and navigating production dynamics at scale. These contributions advance the scientific rigor and understanding of LLMs in social applications serving millions of users.</p> - https://arxiv.org/abs/2603.01973 - Mon, 02 Mar 2026 15:27:31 +0000 + Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards + https://arxiv.org/abs/2603.09117 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09117.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhengzhao Ma, Xueru Wen, Boxi Cao, Yaojie Lu, Hongyu Lin, Jinglin Yang, Min He, Xianpei Han, Le Sun</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Reinforcement Learning from Verifiable Rewards (RLVR) significantly enhances large language models (LLMs) reasoning but severely suffers from calibration degeneration, where models become excessively over-confident in incorrect answers. Previous studies devote to directly incorporating calibration objective into existing optimization target. However, our theoretical analysis demonstrates that there exists a fundamental gradient conflict between the optimization for maximizing policy accuracy and minimizing calibration error. Building on this insight, we propose DCPO, a simple yet effective framework that systematically decouples reasoning and calibration objectives. Extensive experiments demonstrate that our DCPO not only preserves accuracy on par with GRPO but also achieves the best calibration performance and substantially mitigates the over-confidence issue. Our study provides valuable insights and practical solution for more reliable LLM deployment.</p> + https://arxiv.org/abs/2603.09117 + Tue, 10 Mar 2026 02:47:59 +0000 - MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning - https://arxiv.org/abs/2603.02024 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02024.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiachun Li, Shaoping Huang, Zhuoran Jin, Chenlong Zhang, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao</p><p><b>Upvotes:</b> 42</p><p><b>Summary:</b> Recent progress in the reasoning capabilities of multimodal large language models (MLLMs) has empowered them to address more complex tasks such as scientific analysis and mathematical reasoning. Despite their promise, MLLMs' reasoning abilities across different scenarios in real life remain largely unexplored and lack standardized benchmarks for evaluation. To address this gap, we introduce MMR-Life, a comprehensive benchmark designed to evaluate the diverse multimodal multi-image reasoning capabilities of MLLMs across real-life scenarios. MMR-Life consists of 2,646 multiple-choice questions based on 19,108 images primarily sourced from real-world contexts, comprehensively covering seven reasoning types: abductive, analogical, causal, deductive, inductive, spatial, and temporal. Unlike existing reasoning benchmarks, MMR-Life does not rely on domain-specific expertise but instead requires models to integrate information across multiple images and apply diverse reasoning abilities. The evaluation of 37 advanced models highlights the substantial challenge posed by MMR-Life. Even top models like GPT-5 achieve only 58% accuracy and display considerable variance in performance across reasoning types. Moreover, we analyze the reasoning paradigms of existing MLLMs, exploring how factors such as thinking length, reasoning method, and reasoning type affect their performance. In summary, MMR-Life establishes a comprehensive foundation for evaluating, analyzing, and improving the next generation of multimodal reasoning systems.</p> - https://arxiv.org/abs/2603.02024 - Mon, 02 Mar 2026 16:06:23 +0000 + RubiCap: Rubric-Guided Reinforcement Learning for Dense Image Captioning + https://arxiv.org/abs/2603.09160 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09160.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tzu-Heng Huang, Sirajul Salekin, Javier Movellan, Frederic Sala, Manjot Bilkhu</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Dense image captioning is critical for cross-modal alignment in vision-language pretraining and text-to-image generation, but scaling expert-quality annotations is prohibitively expensive. While synthetic captioning via strong vision-language models (VLMs) is a practical alternative, supervised distillation often yields limited output diversity and weak generalization. Reinforcement learning (RL) could overcome these limitations, but its successes have so far been concentrated in verifiable domains that rely on deterministic checkers -- a luxury not available in open-ended captioning. We address this bottleneck with RubiCap, a novel RL framework that derives fine-grained, sample-specific reward signals from LLM-written rubrics. RubiCap first assembles a diverse committee of candidate captions, then employs an LLM rubric writer to extract consensus strengths and diagnose deficiencies in the current policy. These insights are converted into explicit evaluation criteria, enabling an LLM judge to decompose holistic quality assessment and replace coarse scalar rewards with structured, multi-faceted evaluations. Across extensive benchmarks, RubiCap achieves the highest win rates on CapArena, outperforming supervised distillation, prior RL methods, human-expert annotations, and GPT-4V-augmented outputs. On CaptionQA, it demonstrates superior word efficiency: our 7B model matches Qwen2.5-VL-32B-Instruct, and our 3B model surpasses its 7B counterpart. Remarkably, using the compact RubiCap-3B as a captioner produces stronger pretrained VLMs than those trained on captions from proprietary models.</p> + https://arxiv.org/abs/2603.09160 + Tue, 10 Mar 2026 03:51:27 +0000 - WorldStereo: Bridging Camera-Guided Video Generation and Scene Reconstruction via 3D Geometric Memories - https://arxiv.org/abs/2603.02049 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02049.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yisu Zhang, Chenjie Cao, Tengfei Wang, Xuhui Zuo, Junta Wu, Jianke Zhu, Chunchao Guo</p><p><b>Upvotes:</b> 16</p><p><b>Summary:</b> Recent advances in foundational Video Diffusion Models (VDMs) have yielded significant progress. Yet, despite the remarkable visual quality of generated videos, reconstructing consistent 3D scenes from these outputs remains challenging, due to limited camera controllability and inconsistent generated content when viewed from distinct camera trajectories. In this paper, we propose WorldStereo, a novel framework that bridges camera-guided video generation and 3D reconstruction via two dedicated geometric memory modules. Formally, the global-geometric memory enables precise camera control while injecting coarse structural priors through incrementally updated point clouds. Moreover, the spatial-stereo memory constrains the model's attention receptive fields with 3D correspondence to focus on fine-grained details from the memory bank. These components enable WorldStereo to generate multi-view-consistent videos under precise camera control, facilitating high-quality 3D reconstruction. Furthermore, the flexible control branch-based WorldStereo shows impressive efficiency, benefiting from the distribution matching distilled VDM backbone without joint training. Extensive experiments across both camera-guided video generation and 3D reconstruction benchmarks demonstrate the effectiveness of our approach. Notably, we show that WorldStereo acts as a powerful world model, tackling diverse scene generation tasks (whether starting from perspective or panoramic images) with high-fidelity 3D results. Models will be released.</p> - https://arxiv.org/abs/2603.02049 - Mon, 02 Mar 2026 16:36:56 +0000 + The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness + https://arxiv.org/abs/2603.09200 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09200.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Subramanyam Sahoo, Aman Chadha, Vinija Jain, Divya Chaudhary</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Situational awareness, the capacity of an AI system to recognize its own nature, understand its training and deployment context, and reason strategically about its circumstances, is widely considered among the most dangerous emergent capabilities in advanced AI systems. Separately, a growing research effort seeks to improve the logical reasoning capabilities of large language models (LLMs) across deduction, induction, and abduction. In this paper, we argue that these two research trajectories are on a collision course. We introduce the RAISE framework (Reasoning Advancing Into Self Examination), which identifies three mechanistic pathways through which improvements in logical reasoning enable progressively deeper levels of situational awareness: deductive self inference, inductive context recognition, and abductive self modeling. We formalize each pathway, construct an escalation ladder from basic self recognition to strategic deception, and demonstrate that every major research topic in LLM logical reasoning maps directly onto a specific amplifier of situational awareness. We further analyze why current safety measures are insufficient to prevent this escalation. We conclude by proposing concrete safeguards, including a "Mirror Test" benchmark and a Reasoning Safety Parity Principle, and pose an uncomfortable but necessary question to the logical reasoning community about its responsibility in this trajectory.</p> + https://arxiv.org/abs/2603.09200 + Tue, 10 Mar 2026 05:18:48 +0000 - Recursive Think-Answer Process for LLMs and VLMs - https://arxiv.org/abs/2603.02099 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02099.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Byung-Kwan Lee, Youngchae Chee, Yong Man Ro</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Think-Answer reasoners such as DeepSeek-R1 have made notable progress by leveraging interpretable internal reasoning. However, despite the frequent presence of self-reflective cues like "Oops!", they remain vulnerable to output errors during single-pass inference. To address this limitation, we propose an efficient Recursive Think-Answer Process (R-TAP) that enables models to engage in iterative reasoning cycles and generate more accurate answers, going beyond conventional single-pass approaches. Central to this approach is a confidence generator that evaluates the certainty of model responses and guides subsequent improvements. By incorporating two complementary rewards-Recursively Confidence Increase Reward and Final Answer Confidence Reward-we show that R-TAP-enhanced models consistently outperform conventional single-pass methods for both large language models (LLMs) and vision-language models (VLMs). Moreover, by analyzing the frequency of "Oops"-like expressions in model responses, we find that R-TAP-applied models exhibit significantly fewer self-reflective patterns, resulting in more stable and faster inference-time reasoning. We hope R-TAP pave the way evolving into efficient and elaborated methods to refine the reasoning processes of future AI.</p> - https://arxiv.org/abs/2603.02099 - Mon, 02 Mar 2026 17:20:10 +0000 + MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data + https://arxiv.org/abs/2603.09206 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09206.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zongxia Li, Hongyang Du, Chengsong Huang, Xiyang Wu, Lantao Yu, Yicheng He, Jing Xie, Xiaomin Wu, Zhichao Liu, Jiarui Zhang, Fuxiao Liu</p><p><b>Upvotes:</b> 43</p><p><b>Summary:</b> Self-evolving has emerged as a key paradigm for improving foundational models such as Large Language Models (LLMs) and Vision Language Models (VLMs) with minimal human intervention. While recent approaches have demonstrated that LLM agents can self-evolve from scratch with little to no data, VLMs introduce an additional visual modality that typically requires at least some seed data, such as images, to bootstrap the self-evolution process. In this work, we present Multi-model Multimodal Zero (MM-Zero), the first RL-based framework to achieve zero-data self-evolution for VLM reasoning. Moving beyond prior dual-role (Proposer and Solver) setups, MM-Zero introduces a multi-role self-evolving training framework comprising three specialized roles: a Proposer that generates abstract visual concepts and formulates questions; a Coder that translates these concepts into executable code (e.g., Python, SVG) to render visual images; and a Solver that performs multimodal reasoning over the generated visual content. All three roles are initialized from the same base model and trained using Group Relative Policy Optimization (GRPO), with carefully designed reward mechanisms that integrate execution feedback, visual verification, and difficulty balancing. Our experiments show that MM-Zero improves VLM reasoning performance across a wide range of multimodal benchmarks. MM-Zero establishes a scalable path toward self-evolving multi-model systems for multimodal models, extending the frontier of self-improvement beyond the conventional two-model paradigm.</p> + https://arxiv.org/abs/2603.09206 + Tue, 10 Mar 2026 05:23:26 +0000 - OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens - https://arxiv.org/abs/2603.02138 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02138.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yiying Yang, Wei Cheng, Sijin Chen, Honghao Fu, Xianfang Zeng, Yujun Cai, Gang Yu, Xingjun Ma</p><p><b>Upvotes:</b> 134</p><p><b>Summary:</b> OmniLottie is a versatile framework that generates high quality vector animations from multi-modal instructions. For flexible motion and visual content control, we focus on Lottie, a light weight JSON formatting for both shapes and animation behaviors representation. However, the raw Lottie JSON files contain extensive invariant structural metadata and formatting tokens, posing significant challenges for learning vector animation generation. Therefore, we introduce a well designed Lottie tokenizer that transforms JSON files into structured sequences of commands and parameters representing shapes, animation functions and control parameters. Such tokenizer enables us to build OmniLottie upon pretrained vision language models to follow multi-modal interleaved instructions and generate high quality vector animations. To further advance research in vector animation generation, we curate MMLottie-2M, a large scale dataset of professionally designed vector animations paired with textual and visual annotations. With extensive experiments, we validate that OmniLottie can produce vivid and semantically aligned vector animations that adhere closely to multi modal human instructions.</p> - https://arxiv.org/abs/2603.02138 - Mon, 02 Mar 2026 17:59:05 +0000 + Flash-KMeans: Fast and Memory-Efficient Exact K-Means + https://arxiv.org/abs/2603.09229 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09229.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shuo Yang, Haocheng Xi, Yilong Zhao, Muyang Li, Xiaoze Fan, Jintao Zhang, Han Cai, Yujun Lin, Xiuyu Li, Kurt Keutzer, Song Han, Chenfeng Xu, Ion Stoica</p><p><b>Upvotes:</b> 65</p><p><b>Summary:</b> k-means has historically been positioned primarily as an offline processing primitive, typically used for dataset organization or embedding preprocessing rather than as a first-class component in online systems. In this work, we revisit this classical algorithm under the lens of modern AI system design and enable k-means as an online primitive. We point out that existing GPU implementations of k-means remain fundamentally bottlenecked by low-level system constraints rather than theoretical algorithmic complexity. Specifically, the assignment stage suffers from a severe IO bottleneck due to the massive explicit materialization of the N times K distance matrix in High Bandwidth Memory (HBM). Simultaneously, the centroid update stage is heavily penalized by hardware-level atomic write contention caused by irregular, scatter-style token aggregations. To bridge this performance gap, we propose flash-kmeans, an IO-aware and contention-free k-means implementation for modern GPU workloads. Flash-kmeans introduces two core kernel-level innovations: (1) FlashAssign, which fuses distance computation with an online argmin to completely bypass intermediate memory materialization; (2) sort-inverse update, which explicitly constructs an inverse mapping to transform high-contention atomic scatters into high-bandwidth, segment-level localized reductions. Furthermore, we integrate algorithm-system co-designs, including chunked-stream overlap and cache-aware compile heuristics, to ensure practical deployability. Extensive evaluations on NVIDIA H200 GPUs demonstrate that flash-kmeans achieves up to 17.9times end-to-end speedup over best baselines, while outperforming industry-standard libraries like cuML and FAISS by 33times and over 200times, respectively.</p> + https://arxiv.org/abs/2603.09229 + Tue, 10 Mar 2026 05:54:52 +0000 - SageBwd: A Trainable Low-bit Attention - https://arxiv.org/abs/2603.02170 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02170.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jintao Zhang, Marco Chen, Haoxu Wang, Kai Jiang, Ion Stoica, Joseph E. Gonzalez, Jianfei Chen, Jun Zhu</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Low-bit attention, such as SageAttention, has emerged as an effective approach for accelerating model inference, but its applicability to training remains poorly understood. In prior work, we introduced SageBwd, a trainable INT8 attention that quantizes six of seven attention matrix multiplications while preserving fine-tuning performance. However, SageBwd exhibited a persistent performance gap to full-precision attention (FPA) during pre-training. In this work, we investigate why this gap occurs and demonstrate that SageBwd matches full-precision attention during pretraining. Through experiments and theoretical analysis, we reach a few important insights and conclusions: (i) QK-norm is necessary for stable training at large tokens per step, (ii) quantization errors primarily arise from the backward-pass score gradient dS, (iii) reducing tokens per step enables SageBwd to match FPA performance in pre-training, and (iv) K-smoothing remains essential for training stability, while Q-smoothing provides limited benefit during pre-training.</p> - https://arxiv.org/abs/2603.02170 - Mon, 02 Mar 2026 18:39:49 +0000 + Streaming Autoregressive Video Generation via Diagonal Distillation + https://arxiv.org/abs/2603.09488 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09488.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jinxiu Liu, Xuanming Liu, Kangfu Mei, Yandong Wen, Ming-HsuanYang, Weiyang Liu</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Large pretrained diffusion models have significantly enhanced the quality of generated videos, and yet their use in real-time streaming remains limited. Autoregressive models offer a natural framework for sequential frame synthesis but require heavy computation to achieve high fidelity. Diffusion distillation can compress these models into efficient few-step variants, but existing video distillation approaches largely adapt image-specific methods that neglect temporal dependencies. These techniques often excel in image generation but underperform in video synthesis, exhibiting reduced motion coherence, error accumulation over long sequences, and a latency-quality trade-off. We identify two factors that result in these limitations: insufficient utilization of temporal context during step reduction and implicit prediction of subsequent noise levels in next-chunk prediction (i.e., exposure bias). To address these issues, we propose Diagonal Distillation, which operates orthogonally to existing approaches and better exploits temporal information across both video chunks and denoising steps. Central to our approach is an asymmetric generation strategy: more steps early, fewer steps later. This design allows later chunks to inherit rich appearance information from thoroughly processed early chunks, while using partially denoised chunks as conditional inputs for subsequent synthesis. By aligning the implicit prediction of subsequent noise levels during chunk generation with the actual inference conditions, our approach mitigates error propagation and reduces oversaturation in long-range sequences. We further incorporate implicit optical flow modeling to preserve motion quality under strict step constraints. Our method generates a 5-second video in 2.61 seconds (up to 31 FPS), achieving a 277.3x speedup over the undistilled model.</p> + https://arxiv.org/abs/2603.09488 + Tue, 10 Mar 2026 10:45:24 +0000 - Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance - https://arxiv.org/abs/2603.02175 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02175.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yiqi Lin, Guoqiang Liang, Ziyun Zeng, Zechen Bai, Yanzhe Chen, Mike Zheng Shou</p><p><b>Upvotes:</b> 19</p><p><b>Summary:</b> Instruction-based video editing has witnessed rapid progress, yet current methods often struggle with precise visual control, as natural language is inherently limited in describing complex visual nuances. Although reference-guided editing offers a robust solution, its potential is currently bottlenecked by the scarcity of high-quality paired training data. To bridge this gap, we introduce a scalable data generation pipeline that transforms existing video editing pairs into high-fidelity training quadruplets, leveraging image generative models to create synthesized reference scaffolds. Using this pipeline, we construct RefVIE, a large-scale dataset tailored for instruction-reference-following tasks, and establish RefVIE-Bench for comprehensive evaluation. Furthermore, we propose a unified editing architecture, Kiwi-Edit, that synergizes learnable queries and latent visual features for reference semantic guidance. Our model achieves significant gains in instruction following and reference fidelity via a progressive multi-stage training curriculum. Extensive experiments demonstrate that our data and architecture establish a new state-of-the-art in controllable video editing. All datasets, models, and code is released at https://github.com/showlab/Kiwi-Edit.</p> - https://arxiv.org/abs/2603.02175 - Mon, 02 Mar 2026 18:46:28 +0000 + MiniAppBench: Evaluating the Shift from Text to Interactive HTML Responses in LLM-Powered Assistants + https://arxiv.org/abs/2603.09652 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09652.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zuhao Zhang, Chengyue Yu, Yuante Li, Chenyi Zhuang, Linjian Mo, Shuai Li</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> With the rapid advancement of Large Language Models (LLMs) in code generation, human-AI interaction is evolving from static text responses to dynamic, interactive HTML-based applications, which we term MiniApps. These applications require models to not only render visual interfaces but also construct customized interaction logic that adheres to real-world principles. However, existing benchmarks primarily focus on algorithmic correctness or static layout reconstruction, failing to capture the capabilities required for this new paradigm. To address this gap, we introduce MiniAppBench, the first comprehensive benchmark designed to evaluate principle-driven, interactive application generation. Sourced from a real-world application with 10M+ generations, MiniAppBench distills 500 tasks across six domains (e.g., Games, Science, and Tools). Furthermore, to tackle the challenge of evaluating open-ended interactions where no single ground truth exists, we propose MiniAppEval, an agentic evaluation framework. Leveraging browser automation, it performs human-like exploratory testing to systematically assess applications across three dimensions: Intention, Static, and Dynamic. Our experiments reveal that current LLMs still face significant challenges in generating high-quality MiniApps, while MiniAppEval demonstrates high alignment with human judgment, establishing a reliable standard for future research. Our code is available in github.com/MiniAppBench.</p> + https://arxiv.org/abs/2603.09652 + Tue, 10 Mar 2026 13:30:03 +0000 - Tool Verification for Test-Time Reinforcement Learning - https://arxiv.org/abs/2603.02203 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02203.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ruotong Liao, Nikolai Röhrich, Xiaohan Wang, Yuhui Zhang, Yasaman Samadzadeh, Volker Tresp, Serena Yeung-Levy</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Test-time reinforcement learning (TTRL) has emerged as a promising paradigm for self-evolving large reasoning models (LRMs), enabling online adaptation on unlabeled test inputs via self-induced rewards through majority voting. However, a spurious yet high-frequency unverified consensus can become a biased and reinforced reward signal, leading to incorrect mode collapse. We address this failure mode with T^3RL (Tool-Verification for Test-Time Reinforcement Learning), which introduces test-time tool verification into reward estimation. Concretely, a verifier uses an external tool as evidence (e.g., from code execution) to upweight verified rollouts in a verification-aware voting, producing more reliable pseudo-labels for training. Across various math difficulties (MATH-500, AMC, and AIME 2024) and diverse backbone types, T^3RL significantly improves over TTRL, with larger gains on harder problems. More broadly, T^3RL can be viewed as verified online data synthesis, highlighting test-time tool verification as a key mechanism for stabilizing self-evolution.</p> - https://arxiv.org/abs/2603.02203 - Mon, 02 Mar 2026 18:57:52 +0000 + RbtAct: Rebuttal as Supervision for Actionable Review Feedback Generation + https://arxiv.org/abs/2603.09723 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09723.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sihong Wu, Yiling Ma, Yilun Zhao, Tiansheng Hu, Owen Jiang, Manasi Patwardhan, Arman Cohan</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Large language models (LLMs) are increasingly used across the scientific workflow, including to draft peer-review reports. However, many AI-generated reviews are superficial and insufficiently actionable, leaving authors without concrete, implementable guidance and motivating the gap this work addresses. We propose RbtAct, which targets actionable review feedback generation and places existing peer review rebuttal at the center of learning. Rebuttals show which reviewer comments led to concrete revisions or specific plans, and which were only defended. Building on this insight, we leverage rebuttal as implicit supervision to directly optimize a feedback generator for actionability. To support this objective, we propose a new task called perspective-conditioned segment-level review feedback generation, in which the model is required to produce a single focused comment based on the complete paper and a specified perspective such as experiments and writing. We also build a large dataset named RMR-75K that maps review segments to the rebuttal segments that address them, with perspective labels and impact categories that order author uptake. We then train the Llama-3.1-8B-Instruct model with supervised fine-tuning on review segments followed by preference optimization using rebuttal derived pairs. Experiments with human experts and LLM-as-a-judge show consistent gains in actionability and specificity over strong baselines while maintaining grounding and relevance.</p> + https://arxiv.org/abs/2603.09723 + Tue, 10 Mar 2026 14:30:55 +0000 - Reasoning Core: A Scalable Procedural Data Generation Suite for Symbolic Pre-training and Post-Training - https://arxiv.org/abs/2603.02208 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02208.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Valentin Lacombe, Valentin Quesnel, Damien Sileo</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Training on verifiable symbolic data is a promising way to expand the reasoning frontier of language models beyond what standard pre-training corpora provide. Yet existing procedural generators often rely on fixed puzzles or templates and do not deliver the distributional breadth needed at scale. We introduce Reasoning Core, a scalable suite that procedurally generates verifiable symbolic reasoning data across core formal domains: PDDL planning over randomized domains, first-order logic with equality, context-free grammar parsing and generation, causal reasoning over random Bayesian networks, and systems of equations. Each task is paired with an external solver for rigorous verification and admits continuous difficulty control for curriculum design. Examples can optionally include solver-derived reasoning traces, enabling supervised training from the earliest pre-training stages, and the same interface provides verifiable reward functions for reinforcement learning. Our experiments show that mixing Reasoning Core data into pre-training improves downstream reasoning while preserving, or slightly improving, language modeling quality. Zero-shot evaluations confirm these tasks challenge frontier models such as GPT-5. The code and data are publicly available under the MIT license.</p> - https://arxiv.org/abs/2603.02208 - Mon, 02 Mar 2026 18:59:29 +0000 + MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents + https://arxiv.org/abs/2603.09827 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09827.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kangsan Kim, Yanlai Yang, Suji Kim, Woongyeong Yeo, Youngwan Lee, Mengye Ren, Sung Ju Hwang</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> As embodied models become powerful, humans will collaborate with multiple embodied AI agents at their workplace or home in the future. To ensure better communication between human users and the multi-agent system, it is crucial to interpret incoming information from agents in parallel and refer to the appropriate context for each query. Existing challenges include effectively compressing and communicating high volumes of individual sensory inputs in the form of video and correctly aggregating multiple egocentric videos to construct system-level memory. In this work, we first formally define a novel problem of understanding multiple long-horizon egocentric videos simultaneously collected from embodied agents. To facilitate research in this direction, we introduce MultiAgent-EgoQA (MA-EgoQA), a benchmark designed to systemically evaluate existing models in our scenario. MA-EgoQA provides 1.7k questions unique to multiple egocentric streams, spanning five categories: social interaction, task coordination, theory-of-mind, temporal reasoning, and environmental interaction. We further propose a simple baseline model for MA-EgoQA named EgoMAS, which leverages shared memory across embodied agents and agent-wise dynamic retrieval. Through comprehensive evaluation across diverse baselines and EgoMAS on MA-EgoQA, we find that current approaches are unable to effectively handle multiple egocentric streams, highlighting the need for future advances in system-level understanding across the agents. The code and benchmark are available at https://ma-egoqa.github.io.</p> + https://arxiv.org/abs/2603.09827 + Tue, 10 Mar 2026 15:48:35 +0000 - HiFi-Inpaint: Towards High-Fidelity Reference-Based Inpainting for Generating Detail-Preserving Human-Product Images - https://arxiv.org/abs/2603.02210 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02210.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yichen Liu, Donghao Zhou, Jie Wang, Xin Gao, Guisheng Liu, Jiatong Li, Quanwei Zhang, Qiang Lyu, Lanqing Guo, Shilei Wen, Weiqiang Wang, Pheng-Ann Heng</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> Human-product images, which showcase the integration of humans and products, play a vital role in advertising, e-commerce, and digital marketing. The essential challenge of generating such images lies in ensuring the high-fidelity preservation of product details. Among existing paradigms, reference-based inpainting offers a targeted solution by leveraging product reference images to guide the inpainting process. However, limitations remain in three key aspects: the lack of diverse large-scale training data, the struggle of current models to focus on product detail preservation, and the inability of coarse supervision for achieving precise guidance. To address these issues, we propose HiFi-Inpaint, a novel high-fidelity reference-based inpainting framework tailored for generating human-product images. HiFi-Inpaint introduces Shared Enhancement Attention (SEA) to refine fine-grained product features and Detail-Aware Loss (DAL) to enforce precise pixel-level supervision using high-frequency maps. Additionally, we construct a new dataset, HP-Image-40K, with samples curated from self-synthesis data and processed with automatic filtering. Experimental results show that HiFi-Inpaint achieves state-of-the-art performance, delivering detail-preserving human-product images.</p> - https://arxiv.org/abs/2603.02210 - Mon, 02 Mar 2026 18:59:36 +0000 + InternVL-U: Democratizing Unified Multimodal Models for Understanding, Reasoning, Generation and Editing + https://arxiv.org/abs/2603.09877 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09877.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Changyao Tian, Danni Yang, Guanzhou Chen, Erfei Cui, Zhaokai Wang, Yuchen Duan, Penghao Yin, Sitao Chen, Ganlin Yang, Mingxin Liu, Zirun Zhu, Ziqian Fan, Leyao Gu, Haomin Wang, Qi Wei, Jinhui Yin, Xue Yang, Zhihang Zhong, Qi Qin, Yi Xin, Bin Fu, Yihao Liu, Jiaye Ge, Qipeng Guo, Gen Luo, Hongsheng Li, Yu Qiao, Kai Chen, Hongjie Zhang</p><p><b>Upvotes:</b> 41</p><p><b>Summary:</b> Unified multimodal models (UMMs) that integrate understanding, reasoning, generation, and editing face inherent trade-offs between maintaining strong semantic comprehension and acquiring powerful generation capabilities. In this report, we present InternVL-U, a lightweight 4B-parameter UMM that democratizes these capabilities within a unified framework. Guided by the principles of unified contextual modeling and modality-specific modular design with decoupled visual representations, InternVL-U integrates a state-of-the-art Multimodal Large Language Model (MLLM) with a specialized MMDiT-based visual generation head. To further bridge the gap between aesthetic generation and high-level intelligence, we construct a comprehensive data synthesis pipeline targeting high-semantic-density tasks, such as text rendering and scientific reasoning, under a reasoning-centric paradigm that leverages Chain-of-Thought (CoT) to better align abstract user intent with fine-grained visual generation details. Extensive experiments demonstrate that InternVL-U achieves a superior performance - efficiency balance. Despite using only 4B parameters, it consistently outperforms unified baseline models with over 3x larger scales such as BAGEL (14B) on various generation and editing tasks, while retaining strong multimodal understanding and reasoning capabilities.</p> + https://arxiv.org/abs/2603.09877 + Tue, 10 Mar 2026 16:38:33 +0000 - PRISM: Pushing the Frontier of Deep Think via Process Reward Model-Guided Inference - https://arxiv.org/abs/2603.02479 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02479.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Rituraj Sharma, Weiyuan Chen, Noah Provenzano, Tu Vu</p><p><b>Upvotes:</b> 18</p><p><b>Summary:</b> DEEPTHINK methods improve reasoning by generating, refining, and aggregating populations of candidate solutions, which enables strong performance on complex mathematical and scientific tasks. However, existing frameworks often lack reliable correctness signals during inference, which creates a population-enhancement bottleneck where deeper deliberation amplifies errors, suppresses correct minority solutions, and yields weak returns to additional compute. In this paper, we introduce a functional decomposition of DEEPTHINK systems and propose PRISM, a Process Reward Model (PRM)-guided inference algorithm that uses step-level verification to guide both population refinement and solution aggregation. During refinement, PRISM treats candidate solutions as particles in a PRM-defined energy landscape and reshapes the population through score-guided resampling and stochastic refinement, which concentrates probability mass on higher-quality reasoning while preserving diversity. Across mathematics and science benchmarks, PRISM is competitive with or outperforms existing DEEPTHINK methods, reaching 90.0%, 75.4%, and 71.4% with gpt-oss-20b on AIME25, HMMT25, and GPQA Diamond, respectively, while matching or exceeding gpt-oss-120b. Additionally, our analysis shows that PRISM produces consistent net-directional correction during refinement, remains reliable when the initial population contains few correct candidates, and often lies on the compute-accuracy Pareto frontier.</p> - https://arxiv.org/abs/2603.02479 - Tue, 03 Mar 2026 00:03:42 +0000 + Stepping VLMs onto the Court: Benchmarking Spatial Intelligence in Sports + https://arxiv.org/abs/2603.09896 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09896.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuchen Yang, Yuqing Shao, Duxiu Huang, Linfeng Dong, Yifei Liu, Suixin Tang, Xiang Zhou, Yuanyuan Gao, Wei Wang, Yue Zhou, Xue Yang, Yanfeng Wang, Xiao Sun, Zhihang Zhong</p><p><b>Upvotes:</b> 25</p><p><b>Summary:</b> Sports have long attracted broad attention as they push the limits of human physical and cognitive capabilities. Amid growing interest in spatial intelligence for vision-language models (VLMs), sports provide a natural testbed for understanding high-intensity human motion and dynamic object interactions. To this end, we present CourtSI, the first large-scale spatial intelligence dataset tailored to sports scenarios. CourtSI contains over 1M QA pairs, organized under a holistic taxonomy that systematically covers spatial counting, distance measurement, localization, and relational reasoning, across representative net sports including badminton, tennis, and table tennis. Leveraging well-defined court geometry as metric anchors, we develop a semi-automatic data engine to reconstruct sports scenes, enabling scalable curation of CourtSI. In addition, we introduce CourtSI-Bench, a high-quality evaluation benchmark comprising 3,686 QA pairs with rigorous human verification. We evaluate 25 proprietary and open-source VLMs on CourtSI-Bench, revealing a remaining human-AI performance gap and limited generalization from existing spatial intelligence benchmarks. These findings indicate that sports scenarios expose limitations in spatial intelligence capabilities captured by existing benchmarks. Further, fine-tuning Qwen3-VL-8B on CourtSI improves accuracy on CourtSI-Bench by 23.5 percentage points. The adapted model also generalizes effectively to CourtSI-Ext, an evaluation set built on a similar but unseen sport, and demonstrates enhanced spatial-aware commentary generation. Together, these findings demonstrate that CourtSI provides a scalable pathway toward advancing spatial intelligence of VLMs in sports.</p> + https://arxiv.org/abs/2603.09896 + Tue, 10 Mar 2026 16:50:32 +0000 - MUSE: A Run-Centric Platform for Multimodal Unified Safety Evaluation of Large Language Models - https://arxiv.org/abs/2603.02482 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02482.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhongxi Wang, Yueqian Lin, Jingyang Zhang, Hai Helen Li, Yiran Chen</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Safety evaluation and red-teaming of large language models remain predominantly text-centric, and existing frameworks lack the infrastructure to systematically test whether alignment generalizes to audio, image, and video inputs. We present MUSE (Multimodal Unified Safety Evaluation), an open-source, run-centric platform that integrates automatic cross-modal payload generation, three multi-turn attack algorithms (Crescendo, PAIR, Violent Durian), provider-agnostic model routing, and an LLM judge with a five-level safety taxonomy into a single browser-based system. A dual-metric framework distinguishes hard Attack Success Rate (Compliance only) from soft ASR (including Partial Compliance), capturing partial information leakage that binary metrics miss. To probe whether alignment generalizes across modality boundaries, we introduce Inter-Turn Modality Switching (ITMS), which augments multi-turn attacks with per-turn modality rotation. Experiments across six multimodal LLMs from four providers show that multi-turn strategies can achieve up to 90-100% ASR against models with near-perfect single-turn refusal. ITMS does not uniformly raise final ASR on already-saturated baselines, but accelerates convergence by destabilizing early-turn defenses, and ablation reveals that the direction of modality effects is model-family-specific rather than universal, underscoring the need for provider-aware cross-modal safety testing.</p> - https://arxiv.org/abs/2603.02482 - Tue, 03 Mar 2026 00:10:23 +0000 + Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs + https://arxiv.org/abs/2603.09906 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09906.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zorik Gekhman, Roee Aharoni, Eran Ofek, Mor Geva, Roi Reichart, Jonathan Herzig</p><p><b>Upvotes:</b> 61</p><p><b>Summary:</b> While reasoning in LLMs plays a natural role in math, code generation, and multi-hop factual questions, its effect on simple, single-hop factual questions remains unclear. Such questions do not require step-by-step logical decomposition, making the utility of reasoning highly counterintuitive. Nevertheless, we find that enabling reasoning substantially expands the capability boundary of the model's parametric knowledge recall, unlocking correct answers that are otherwise effectively unreachable. Why does reasoning aid parametric knowledge recall when there are no complex reasoning steps to be done? To answer this, we design a series of hypothesis-driven controlled experiments, and identify two key driving mechanisms: (1) a computational buffer effect, where the model uses the generated reasoning tokens to perform latent computation independent of their semantic content; and (2) factual priming, where generating topically related facts acts as a semantic bridge that facilitates correct answer retrieval. Importantly, this latter generative self-retrieval mechanism carries inherent risks: we demonstrate that hallucinating intermediate facts during reasoning increases the likelihood of hallucinations in the final answer. Finally, we show that our insights can be harnessed to directly improve model accuracy by prioritizing reasoning trajectories that contain hallucination-free factual statements.</p> + https://arxiv.org/abs/2603.09906 + Tue, 10 Mar 2026 16:59:20 +0000 - MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning - https://arxiv.org/abs/2603.03379 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03379.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiejun Tan, Zhicheng Dou, Liancheng Zhang, Yuyang Hu, Yiruo Cheng, Ji-Rong Wen</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> As Large Language Models (LLMs) are increasingly used for long-duration tasks, maintaining effective long-term memory has become a critical challenge. Current methods often face a trade-off between cost and accuracy. Simple storage methods often fail to retrieve relevant information, while complex indexing methods (such as memory graphs) require heavy computation and can cause information loss. Furthermore, relying on the working LLM to process all memories is computationally expensive and slow. To address these limitations, we propose MemSifter, a novel framework that offloads the memory retrieval process to a small-scale proxy model. Instead of increasing the burden on the primary working LLM, MemSifter uses a smaller model to reason about the task before retrieving the necessary information. This approach requires no heavy computation during the indexing phase and adds minimal overhead during inference. To optimize the proxy model, we introduce a memory-specific Reinforcement Learning (RL) training paradigm. We design a task-outcome-oriented reward based on the working LLM's actual performance in completing the task. The reward measures the actual contribution of retrieved memories by mutiple interactions with the working LLM, and discriminates retrieved rankings by stepped decreasing contributions. Additionally, we employ training techniques such as Curriculum Learning and Model Merging to improve performance. We evaluated MemSifter on eight LLM memory benchmarks, including Deep Research tasks. The results demonstrate that our method meets or exceeds the performance of existing state-of-the-art approaches in both retrieval accuracy and final task completion. MemSifter offers an efficient and scalable solution for long-term LLM memory. We have open-sourced the model weights, code, and training data to support further research.</p> - https://arxiv.org/abs/2603.03379 - Tue, 03 Mar 2026 02:57:38 +0000 + Towards a Neural Debugger for Python + https://arxiv.org/abs/2603.09951 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09951.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Maximilian Beck, Jonas Gehring, Jannik Kossen, Gabriel Synnaeve</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Training large language models (LLMs) on Python execution traces grounds them in code execution and enables the line-by-line execution prediction of whole Python programs, effectively turning them into neural interpreters (FAIR CodeGen Team et al., 2025). However, developers rarely execute programs step by step; instead, they use debuggers to stop execution at certain breakpoints and step through relevant portions only while inspecting or modifying program variables. Existing neural interpreter approaches lack such interactive control. To address this limitation, we introduce neural debuggers: language models that emulate traditional debuggers, supporting operations such as stepping into, over, or out of functions, as well as setting breakpoints at specific source lines. We show that neural debuggers -- obtained via fine-tuning large LLMs or pre-training smaller models from scratch -- can reliably model both forward execution (predicting future states and outputs) and inverse execution (inferring prior states or inputs) conditioned on debugger actions. Evaluated on CruxEval, our models achieve strong performance on both output and input prediction tasks, demonstrating robust conditional execution modeling. Our work takes first steps towards future agentic coding systems in which neural debuggers serve as a world model for simulated debugging environments, providing execution feedback or enabling agents to interact with real debugging tools. This capability lays the foundation for more powerful code generation, program understanding, and automated debugging.</p> + https://arxiv.org/abs/2603.09951 + Tue, 10 Mar 2026 17:47:05 +0000 - Track4World: Feedforward World-centric Dense 3D Tracking of All Pixels - https://arxiv.org/abs/2603.02573 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02573.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiahao Lu, Jiayi Xu, Wenbo Hu, Ruijie Zhu, Chengfeng Zhao, Sai-Kit Yeung, Ying Shan, Yuan Liu</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Estimating the 3D trajectory of every pixel from a monocular video is crucial and promising for a comprehensive understanding of the 3D dynamics of videos. Recent monocular 3D tracking works demonstrate impressive performance, but are limited to either tracking sparse points on the first frame or a slow optimization-based framework for dense tracking. In this paper, we propose a feedforward model, called Track4World, enabling an efficient holistic 3D tracking of every pixel in the world-centric coordinate system. Built on the global 3D scene representation encoded by a VGGT-style ViT, Track4World applies a novel 3D correlation scheme to simultaneously estimate the pixel-wise 2D and 3D dense flow between arbitrary frame pairs. The estimated scene flow, along with the reconstructed 3D geometry, enables subsequent efficient 3D tracking of every pixel of this video. Extensive experiments on multiple benchmarks demonstrate that our approach consistently outperforms existing methods in 2D/3D flow estimation and 3D tracking, highlighting its robustness and scalability for real-world 4D reconstruction tasks.</p> - https://arxiv.org/abs/2603.02573 - Tue, 03 Mar 2026 03:45:43 +0000 + CREATE: Testing LLMs for Associative Creativity + https://arxiv.org/abs/2603.09970 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.09970.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Manya Wadhwa, Tiasa Singha Roy, Harvey Lederman, Junyi Jessy Li, Greg Durrett</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> A key component of creativity is associative reasoning: the ability to draw novel yet meaningful connections between concepts. We introduce CREATE, a benchmark designed to evaluate models' capacity for creative associative reasoning. CREATE requires models to generate sets of paths connecting concepts in a model's parametric knowledge. Paths should have high specificity (distinctiveness and closeness of the concept connection) and high diversity (dissimilarity from other paths), and models are scored more highly if they produce a larger set of strong, diverse paths. This task shares demands of real creativity tasks like hypothesis generation, including an extremely large search space, but enables collection of a sizable benchmark with objective answer grading. Evaluation of frontier models shows that the strongest models achieve higher creative utility than others, with the high multiplicity of answers and complexity of the search making benchmark saturation difficult to achieve. Furthermore, our results illustrate that thinking models are not always more effective on our task, even with high token budgets. Recent approaches for creative prompting give some but limited additional improvement. CREATE provides a sandbox for developing new methods to improve models' capacity for associative creativity.</p> + https://arxiv.org/abs/2603.09970 + Tue, 10 Mar 2026 17:58:44 +0000 - How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities - https://arxiv.org/abs/2603.02578 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02578.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ziwen Xu, Kewei Xu, Haoming Xu, Haiwen Hong, Longtao Huang, Hui Xue, Ningyu Zhang, Yongliang Shen, Guozhou Zheng, Huajun Chen, Shumin Deng</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerEval, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.</p> - https://arxiv.org/abs/2603.02578 - Tue, 03 Mar 2026 03:50:13 +0000 + Lost in Backpropagation: The LM Head is a Gradient Bottleneck + https://arxiv.org/abs/2603.10145 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10145.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Nathan Godey, Yoav Artzi</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> The last layer of neural language models (LMs) projects output features of dimension D to logits in dimension V, the size of the vocabulary, where usually D ll V. This mismatch is known to raise risks of limited expressivity in neural LMs, creating a so-called softmax bottleneck. We show the softmax bottleneck is not only an expressivity bottleneck but also an optimization bottleneck. Backpropagating V-dimensional gradients through a rank-D linear layer induces unavoidable compression, which alters the training feedback provided to the vast majority of the parameters. We present a theoretical analysis of this phenomenon and measure empirically that 95-99% of the gradient norm is suppressed by the output layer, resulting in vastly suboptimal update directions. We conduct controlled pretraining experiments showing that the gradient bottleneck makes trivial patterns unlearnable, and drastically affects the training dynamics of LLMs. We argue that this inherent flaw contributes to training inefficiencies at scale independently of the model architecture, and raises the need for new LM head designs.</p> + https://arxiv.org/abs/2603.10145 + Tue, 10 Mar 2026 18:27:01 +0000 - Heterogeneous Agent Collaborative Reinforcement Learning - https://arxiv.org/abs/2603.02604 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02604.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhixia Zhang, Zixuan Huang, Xin Xia, Deqing Wang, Fuzhen Zhuang, Shuai Ma, Ning Ding, Yaodong Yang, Jianxin Li, Yikun Ban</p><p><b>Upvotes:</b> 146</p><p><b>Summary:</b> We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new learning paradigm that addresses the inefficiencies of isolated on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional teacher-to-student transfer. Building on this paradigm, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation and optimization correctness. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO by an average of 3.3\% while using only half the rollout cost.</p> - https://arxiv.org/abs/2603.02604 - Tue, 03 Mar 2026 05:09:49 +0000 + ReMix: Reinforcement routing for mixtures of LoRAs in LLM finetuning + https://arxiv.org/abs/2603.10160 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10160.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ruizhong Qiu, Hanqing Zeng, Yinglong Xia, Yiwen Meng, Ren Chen, Jiarui Feng, Dongqi Fu, Qifan Wang, Jiayi Liu, Jun Xiao, Xiangjun Fan, Benyu Zhang, Hong Li, Zhining Liu, Hyunsik Yoo, Zhichen Zeng, Tianxin Wei, Hanghang Tong</p><p><b>Upvotes:</b> 20</p><p><b>Summary:</b> Low-rank adapters (LoRAs) are a parameter-efficient finetuning technique that injects trainable low-rank matrices into pretrained models to adapt them to new tasks. Mixture-of-LoRAs models expand neural networks efficiently by routing each layer input to a small subset of specialized LoRAs of the layer. Existing Mixture-of-LoRAs routers assign a learned routing weight to each LoRA to enable end-to-end training of the router. Despite their empirical promise, we observe that the routing weights are typically extremely imbalanced across LoRAs in practice, where only one or two LoRAs often dominate the routing weights. This essentially limits the number of effective LoRAs and thus severely hinders the expressive power of existing Mixture-of-LoRAs models. In this work, we attribute this weakness to the nature of learnable routing weights and rethink the fundamental design of the router. To address this critical issue, we propose a new router designed that we call Reinforcement Routing for Mixture-of-LoRAs (ReMix). Our key idea is using non-learnable routing weights to ensure all active LoRAs to be equally effective, with no LoRA dominating the routing weights. However, our routers cannot be trained directly via gradient descent due to our non-learnable routing weights. Hence, we further propose an unbiased gradient estimator for the router by employing the reinforce leave-one-out (RLOO) technique, where we regard the supervision loss as the reward and the router as the policy in reinforcement learning. Our gradient estimator also enables to scale up training compute to boost the predictive performance of our ReMix. Extensive experiments demonstrate that our proposed ReMix significantly outperform state-of-the-art parameter-efficient finetuning methods under a comparable number of activated parameters.</p> + https://arxiv.org/abs/2603.10160 + Tue, 10 Mar 2026 18:51:27 +0000 - DREAM: Where Visual Understanding Meets Text-to-Image Generation - https://arxiv.org/abs/2603.02667 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02667.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chao Li, Tianhong Li, Sai Vidyaranya Nuthalapati, Hong-You Chen, Satya Narayan Shukla, Yonghuan Yang, Jun Xiao, Xiangjun Fan, Aashu Singh, Dina Katabi, Shlok Kumar Mishra</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Unifying visual representation learning and text-to-image (T2I) generation within a single model remains a central challenge in multimodal learning. We introduce DREAM, a unified framework that jointly optimizes discriminative and generative objectives, while learning strong visual representations. DREAM is built on two key techniques: During training, Masking Warmup, a progressive masking schedule, begins with minimal masking to establish the contrastive alignment necessary for representation learning, then gradually transitions to full masking for stable generative training. At inference, DREAM employs Semantically Aligned Decoding to align partially masked image candidates with the target text and select the best one for further decoding, improving text-image fidelity (+6.3%) without external rerankers. Trained solely on CC12M, DREAM achieves 72.7% ImageNet linear-probing accuracy (+1.1% over CLIP) and an FID of 4.25 (+6.2% over FLUID), with consistent gains in few-shot classification, semantic segmentation, and depth estimation. These results demonstrate that discriminative and generative objectives can be synergistic, allowing unified multimodal models that excel at both visual understanding and generation.</p> - https://arxiv.org/abs/2603.02667 - Tue, 03 Mar 2026 06:54:19 +0000 + OpenClaw-RL: Train Any Agent Simply by Talking + https://arxiv.org/abs/2603.10165 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10165.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yinjie Wang, Xuyang Chen, Xiaolong Jin, Mengdi Wang, Ling Yang</p><p><b>Upvotes:</b> 94</p><p><b>Summary:</b> Every agent interaction generates a next-state signal, namely the user reply, tool output, terminal or GUI state change that follows each action, yet no existing agentic RL system recovers it as a live, online learning source. We present OpenClaw-RL, a framework built on a simple observation: next-state signals are universal, and policy can learn from all of them simultaneously. Personal conversations, terminal executions, GUI interactions, SWE tasks, and tool-call traces are not separate training problems. They are all interactions that can be used to train the same policy in the same loop. Next-state signals encode two forms of information: evaluative signals, which indicate how well the action performed and are extracted as scalar rewards via a PRM judge; and directive signals, which indicate how the action should have been different and are recovered through Hindsight-Guided On-Policy Distillation (OPD). We extract textual hints from the next state, construct an enhanced teacher context, and provide token-level directional advantage supervision that is richer than any scalar reward. Due to the asynchronous design, the model serves live requests, the PRM judges ongoing interactions, and the trainer updates the policy at the same time, with zero coordination overhead between them. Applied to personal agents, OpenClaw-RL enables an agent to improve simply by being used, recovering conversational signals from user re-queries, corrections, and explicit feedback. Applied to general agents, the same infrastructure supports scalable RL across terminal, GUI, SWE, and tool-call settings, where we additionally demonstrate the utility of process rewards. Code: https://github.com/Gen-Verse/OpenClaw-RL</p> + https://arxiv.org/abs/2603.10165 + Tue, 10 Mar 2026 18:59:01 +0000 - Next Embedding Prediction Makes World Models Stronger - https://arxiv.org/abs/2603.02765 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02765.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> George Bredis, Nikita Balagansky, Daniil Gavrilov, Ruslan Rakhimov</p><p><b>Upvotes:</b> 17</p><p><b>Summary:</b> Capturing temporal dependencies is critical for model-based reinforcement learning (MBRL) in partially observable, high-dimensional domains. We introduce NE-Dreamer, a decoder-free MBRL agent that leverages a temporal transformer to predict next-step encoder embeddings from latent state sequences, directly optimizing temporal predictive alignment in representation space. This approach enables NE-Dreamer to learn coherent, predictive state representations without reconstruction losses or auxiliary supervision. On the DeepMind Control Suite, NE-Dreamer matches or exceeds the performance of DreamerV3 and leading decoder-free agents. On a challenging subset of DMLab tasks involving memory and spatial reasoning, NE-Dreamer achieves substantial gains. These results establish next-embedding prediction with temporal transformers as an effective, scalable framework for MBRL in complex, partially observable environments.</p> - https://arxiv.org/abs/2603.02765 - Tue, 03 Mar 2026 09:04:28 +0000 + Video-Based Reward Modeling for Computer-Use Agents + https://arxiv.org/abs/2603.10178 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10178.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Linxin Song, Jieyu Zhang, Huanxin Sheng, Taiwei Shi, Gupta Rahul, Yang Liu, Ranjay Krishna, Jian Kang, Jieyu Zhao</p><p><b>Upvotes:</b> 35</p><p><b>Summary:</b> Computer-using agents (CUAs) are becoming increasingly capable; however, it remains difficult to scale evaluation of whether a trajectory truly fulfills a user instruction. In this work, we study reward modeling from execution video: a sequence of keyframes from an agent trajectory that is independent of the agent's internal reasoning or actions. Although video-execution modeling is method-agnostic, it presents key challenges, including highly redundant layouts and subtle, localized cues that determine success. We introduce Execution Video Reward 53k (ExeVR-53k), a dataset of 53k high-quality video--task--reward triplets. We further propose adversarial instruction translation to synthesize negative samples with step-level annotations. To enable learning from long, high-resolution execution videos, we design spatiotemporal token pruning, which removes homogeneous regions and persistent tokens while preserving decisive UI changes. Building on these components, we fine-tune an Execution Video Reward Model (ExeVRM) that takes only a user instruction and a video-execution sequence to predict task success. Our ExeVRM 8B achieves 84.7% accuracy and 87.7% recall on video-execution assessment, outperforming strong proprietary models such as GPT-5.2 and Gemini-3 Pro across Ubuntu, macOS, Windows, and Android, while providing more precise temporal attribution. These results show that video-execution reward modeling can serve as a scalable, model-agnostic evaluator for CUAs.</p> + https://arxiv.org/abs/2603.10178 + Tue, 10 Mar 2026 19:17:22 +0000 - NOVA: Sparse Control, Dense Synthesis for Pair-Free Video Editing - https://arxiv.org/abs/2603.02802 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.02802.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tianlin Pan, Jiayi Dai, Chenpu Yuan, Zhengyao Lv, Binxin Yang, Hubery Yin, Chen Li, Jing Lyu, Caifeng Shan, Chenyang Si</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Recent video editing models have achieved impressive results, but most still require large-scale paired datasets. Collecting such naturally aligned pairs at scale remains highly challenging and constitutes a critical bottleneck, especially for local video editing data. Existing workarounds transfer image editing to video through global motion control for pair-free video editing, but such designs struggle with background and temporal consistency. In this paper, we propose NOVA: Sparse Control \& Dense Synthesis, a new framework for unpaired video editing. Specifically, the sparse branch provides semantic guidance through user-edited keyframes distributed across the video, and the dense branch continuously incorporates motion and texture information from the original video to maintain high fidelity and coherence. Moreover, we introduce a degradation-simulation training strategy that enables the model to learn motion reconstruction and temporal consistency by training on artificially degraded videos, thus eliminating the need for paired data. Our extensive experiments demonstrate that NOVA outperforms existing approaches in edit fidelity, motion preservation, and temporal coherence.</p> - https://arxiv.org/abs/2603.02802 - Tue, 03 Mar 2026 09:41:06 +0000 + DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use + https://arxiv.org/abs/2603.11076 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11076.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Aili Chen, Chi Zhang, Junteng Liu, Jiangjie Chen, Chengyu Du, Yunji Li, Ming Zhong, Qin Wang, Zhengmao Zhu, Jiayuan Song, Ke Ji, Junxian He, Pengyu Zhao, Yanghua Xiao</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Recent work synthesizes agentic tasks for post-training tool-using LLMs, yet robust generalization under shifts in tasks and toolsets remains an open challenge. We trace this brittleness to insufficient diversity in synthesized tasks. Scaling diversity is difficult because training requires tasks to remain executable and verifiable, while generalization demands coverage of diverse tool types, toolset combinations, and heterogeneous tool-use patterns. We propose DIVE, an evidence-driven recipe that inverts synthesis order, executing diverse, real-world tools first and reverse-deriving tasks strictly entailed by the resulting traces, thereby providing grounding by construction. DIVE scales structural diversity along two controllable axes, tool-pool coverage and per-task toolset variety, and an Evidence Collection--Task Derivation loop further induces rich multi-step tool-use patterns across 373 tools in five domains. Training Qwen3-8B on DIVE data (48k SFT + 3.2k RL) improves by +22 average points across 9 OOD benchmarks and outperforms the strongest 8B baseline by +68. Remarkably, controlled scaling analysis reveals that diversity scaling consistently outperforms quantity scaling for OOD generalization, even with 4x less data.</p> + https://arxiv.org/abs/2603.11076 + Tue, 10 Mar 2026 20:54:23 +0000 - Kling-MotionControl Technical Report - https://arxiv.org/abs/2603.03160 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03160.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kling Team, Jialu Chen, Yikang Ding, Zhixue Fang, Kun Gai, Kang He, Xu He, Jingyun Hua, Mingming Lao, Xiaohan Li, Hui Liu, Jiwen Liu, Xiaoqiang Liu, Fan Shi, Xiaoyu Shi, Peiqin Sun, Songlin Tang, Pengfei Wan, Tiancheng Wen, Zhiyong Wu, Haoxian Zhang, Runze Zhao, Yuanxing Zhang, Yan Zhou</p><p><b>Upvotes:</b> 25</p><p><b>Summary:</b> Character animation aims to generate lifelike videos by transferring motion dynamics from a driving video to a reference image. Recent strides in generative models have paved the way for high-fidelity character animation. In this work, we present Kling-MotionControl, a unified DiT-based framework engineered specifically for robust, precise, and expressive holistic character animation. Leveraging a divide-and-conquer strategy within a cohesive system, the model orchestrates heterogeneous motion representations tailored to the distinct characteristics of body, face, and hands, effectively reconciling large-scale structural stability with fine-grained articulatory expressiveness. To ensure robust cross-identity generalization, we incorporate adaptive identity-agnostic learning, facilitating natural motion retargeting for diverse characters ranging from realistic humans to stylized cartoons. Simultaneously, we guarantee faithful appearance preservation through meticulous identity injection and fusion designs, further supported by a subject library mechanism that leverages comprehensive reference contexts. To ensure practical utility, we implement an advanced acceleration framework utilizing multi-stage distillation, boosting inference speed by over 10x. Kling-MotionControl distinguishes itself through intelligent semantic motion understanding and precise text responsiveness, allowing for flexible control beyond visual inputs. Human preference evaluations demonstrate that Kling-MotionControl delivers superior performance compared to leading commercial and open-source solutions, achieving exceptional fidelity in holistic motion control, open domain generalization, and visual quality and coherence. These results establish Kling-MotionControl as a robust solution for high-quality, controllable, and lifelike character animation.</p> - https://arxiv.org/abs/2603.03160 - Tue, 03 Mar 2026 17:02:45 +0000 + ID-LoRA: Identity-Driven Audio-Video Personalization with In-Context LoRA + https://arxiv.org/abs/2603.10256 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10256.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Aviad Dahan, Moran Yanuka, Noa Kraicer, Lior Wolf, Raja Giryes</p><p><b>Upvotes:</b> 16</p><p><b>Summary:</b> Existing video personalization methods preserve visual likeness but treat video and audio separately. Without access to the visual scene, audio models cannot synchronize sounds with on-screen actions; and because classical voice-cloning models condition only on a reference recording, a text prompt cannot redirect speaking style or acoustic environment. We propose ID-LoRA (Identity-Driven In-Context LoRA), which jointly generates a subject's appearance and voice in a single model, letting a text prompt, a reference image, and a short audio clip govern both modalities together. ID-LoRA adapts the LTX-2 joint audio-video diffusion backbone via parameter-efficient In-Context LoRA and, to our knowledge, is the first method to personalize visual appearance and voice in a single generative pass. Two challenges arise. Reference and generation tokens share the same positional-encoding space, making them hard to distinguish; we address this with negative temporal positions, placing reference tokens in a disjoint RoPE region while preserving their internal temporal structure. Speaker characteristics also tend to be diluted during denoising; we introduce identity guidance, a classifier-free guidance variant that amplifies speaker-specific features by contrasting predictions with and without the reference signal. In human preference studies, ID-LoRA is preferred over Kling 2.6 Pro by 73% of annotators for voice similarity and 65% for speaking style. On cross-environment settings, speaker similarity improves by 24% over Kling, with the gap widening as conditions diverge. A preliminary user study further suggests that joint generation provides a useful inductive bias for physically grounded sound synthesis. ID-LoRA achieves these results with only ~3K training pairs on a single GPU. Code, models, and data will be released.</p> + https://arxiv.org/abs/2603.10256 + Tue, 10 Mar 2026 22:23:36 +0000 - BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing? - https://arxiv.org/abs/2603.03194 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03194.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Guoxin Chen, Fanzhe Meng, Jiale Zhao, Minghao Li, Daixuan Cheng, Huatong Song, Jie Chen, Yuzhi Lin, Hui Chen, Xin Zhao, Ruihua Song, Chang Liu, Cheng Chen, Kai Jia, Ji-Rong Wen</p><p><b>Upvotes:</b> 50</p><p><b>Summary:</b> Current benchmarks for code agents primarily assess narrow, repository-specific fixes, overlooking critical real-world challenges such as cross-repository reasoning, domain-specialized problem solving, dependency-driven migration, and full-repository generation. To address this gap, we introduce BeyondSWE, a comprehensive benchmark that broadens existing evaluations along two axes - resolution scope and knowledge scope - using 500 real-world instances across four distinct settings. Experimental results reveal a significant capability gap: even frontier models plateau below 45% success, and no single model performs consistently across task types. To systematically investigate the role of external knowledge, we develop SearchSWE, a framework that integrates deep search with coding abilities. Our experiments show that search augmentation yields inconsistent gains and can in some cases degrade performance, highlighting the difficulty of emulating developer-like workflows that interleave search and reasoning during coding tasks. This work offers both a realistic, challenging evaluation benchmark and a flexible framework to advance research toward more capable code agents.</p> - https://arxiv.org/abs/2603.03194 - Tue, 03 Mar 2026 17:52:01 +0000 + Geometric Autoencoder for Diffusion Models + https://arxiv.org/abs/2603.10365 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10365.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hangyu Liu, Jianyong Wang, Yutao Sun</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Latent diffusion models have established a new state-of-the-art in high-resolution visual generation. Integrating Vision Foundation Model priors improves generative efficiency, yet existing latent designs remain largely heuristic. These approaches often struggle to unify semantic discriminability, reconstruction fidelity, and latent compactness. In this paper, we propose Geometric Autoencoder (GAE), a principled framework that systematically addresses these challenges. By analyzing various alignment paradigms, GAE constructs an optimized low-dimensional semantic supervision target from VFMs to provide guidance for the autoencoder. Furthermore, we leverage latent normalization that replaces the restrictive KL-divergence of standard VAEs, enabling a more stable latent manifold specifically optimized for diffusion learning. To ensure robust reconstruction under high-intensity noise, GAE incorporates a dynamic noise sampling mechanism. Empirically, GAE achieves compelling performance on the ImageNet-1K 256 times 256 benchmark, reaching a gFID of 1.82 at only 80 epochs and 1.31 at 800 epochs without Classifier-Free Guidance, significantly surpassing existing state-of-the-art methods. Beyond generative quality, GAE establishes a superior equilibrium between compression, semantic depth and robust reconstruction stability. These results validate our design considerations, offering a promising paradigm for latent diffusion modeling. Code and models are publicly available at https://github.com/freezing-index/Geometric-Autoencoder-for-Diffusion-Models.</p> + https://arxiv.org/abs/2603.10365 + Wed, 11 Mar 2026 03:29:16 +0000 - Chain of World: World Model Thinking in Latent Motion - https://arxiv.org/abs/2603.03195 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03195.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Fuxiang Yang, Donglin Di, Lulu Tang, Xuancheng Zhang, Lei Fan, Hao Li, Chen Wei, Tonghua Su, Baorui Ma</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Vision-Language-Action (VLA) models are a promising path toward embodied intelligence, yet they often overlook the predictive and temporal-causal structure underlying visual dynamics. World-model VLAs address this by predicting future frames, but waste capacity reconstructing redundant backgrounds. Latent-action VLAs encode frame-to-frame transitions compactly, but lack temporally continuous dynamic modeling and world knowledge. To overcome these limitations, we introduce CoWVLA (Chain-of-World VLA), a new "Chain of World" paradigm that unifies world-model temporal reasoning with a disentangled latent motion representation. First, a pretrained video VAE serves as a latent motion extractor, explicitly factorizing video segments into structure and motion latents. Then, during pre-training, the VLA learns from an instruction and an initial frame to infer a continuous latent motion chain and predict the segment's terminal frame. Finally, during co-fine-tuning, this latent dynamic is aligned with discrete action prediction by jointly modeling sparse keyframes and action sequences in a unified autoregressive decoder. This design preserves the world-model benefits of temporal reasoning and world knowledge while retaining the compactness and interpretability of latent actions, enabling efficient visuomotor learning. Extensive experiments on robotic simulation benchmarks show that CoWVLA outperforms existing world-model and latent-action approaches and achieves moderate computational efficiency, highlighting its potential as a more effective VLA pretraining paradigm. The project website can be found at https://fx-hit.github.io/cowvla-io.</p> - https://arxiv.org/abs/2603.03195 - Tue, 03 Mar 2026 17:52:06 +0000 + FireRedASR2S: A State-of-the-Art Industrial-Grade All-in-One Automatic Speech Recognition System + https://arxiv.org/abs/2603.10420 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10420.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kaituo Xu, Yan Jia, Kai Huang, Junjie Chen, Wenpeng Li, Kun Liu, Feng-Long Xie, Xu Tang, Yao Hu</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> We present FireRedASR2S, a state-of-the-art industrial-grade all-in-one automatic speech recognition (ASR) system. It integrates four modules in a unified pipeline: ASR, Voice Activity Detection (VAD), Spoken Language Identification (LID), and Punctuation Prediction (Punc). All modules achieve SOTA performance on the evaluated benchmarks: FireRedASR2: An ASR module with two variants, FireRedASR2-LLM (8B+ parameters) and FireRedASR2-AED (1B+ parameters), supporting speech and singing transcription for Mandarin, Chinese dialects and accents, English, and code-switching. Compared to FireRedASR, FireRedASR2 delivers improved recognition accuracy and broader dialect and accent coverage. FireRedASR2-LLM achieves 2.89% average CER on 4 public Mandarin benchmarks and 11.55% on 19 public Chinese dialects and accents benchmarks, outperforming competitive baselines including Doubao-ASR, Qwen3-ASR, and Fun-ASR. FireRedVAD: An ultra-lightweight module (0.6M parameters) based on the Deep Feedforward Sequential Memory Network (DFSMN), supporting streaming VAD, non-streaming VAD, and multi-label VAD (mVAD). On the FLEURS-VAD-102 benchmark, it achieves 97.57% frame-level F1 and 99.60% AUC-ROC, outperforming Silero-VAD, TEN-VAD, FunASR-VAD, and WebRTC-VAD. FireRedLID: An Encoder-Decoder LID module supporting 100+ languages and 20+ Chinese dialects and accents. On FLEURS (82 languages), it achieves 97.18% utterance-level accuracy, outperforming Whisper and SpeechBrain. FireRedPunc: A BERT-style punctuation prediction module for Chinese and English. On multi-domain benchmarks, it achieves 78.90% average F1, outperforming FunASR-Punc (62.77%). To advance research in speech processing, we release model weights and code at https://github.com/FireRedTeam/FireRedASR2S.</p> + https://arxiv.org/abs/2603.10420 + Wed, 11 Mar 2026 05:11:06 +0000 - Specificity-aware reinforcement learning for fine-grained open-world classification - https://arxiv.org/abs/2603.03197 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03197.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Samuele Angheben, Davide Berasi, Alessandro Conti, Elisa Ricci, Yiming Wang</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Classifying fine-grained visual concepts under open-world settings, i.e., without a predefined label set, demands models to be both accurate and specific. Recent reasoning Large Multimodal Models (LMMs) exhibit strong visual understanding capability but tend to produce overly generic predictions when performing fine-grained image classification. Our preliminary analysis reveals that models do possess the intrinsic fine-grained domain knowledge. However, promoting more specific predictions (specificity) without compromising correct ones (correctness) remains a non-trivial and understudied challenge. In this work, we investigate how to steer reasoning LMMs toward predictions that are both correct and specific. We propose a novel specificity-aware reinforcement learning framework, SpeciaRL, to fine-tune reasoning LMMs on fine-grained image classification under the open-world setting. SpeciaRL introduces a dynamic, verifier-based reward signal anchored to the best predictions within online rollouts, promoting specificity while respecting the model's capabilities to prevent incorrect predictions. Our out-of-domain experiments show that SpeciaRL delivers the best trade-off between correctness and specificity across extensive fine-grained benchmarks, surpassing existing methods and advancing open-world fine-grained image classification. Code and model are publicly available at https://github.com/s-angheben/SpeciaRL.</p> - https://arxiv.org/abs/2603.03197 - Tue, 03 Mar 2026 17:52:39 +0000 + The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training + https://arxiv.org/abs/2603.10444 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10444.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hengjie Cao, Zhendong Huang, Mengyi Chen, Yifeng Yang, Fanqi Yu, Ruijun Huang, Fang Dong, Xin Zhang, Jixian Zhou, Anrui Chen, Mingzhi Dong, Yujiang Wang, Jinlong Hou, Qin Lv, Yuan Cheng, Tun Lu, Fan Yang, Li Shang</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Large language models trained on natural language exhibit pronounced anisotropy: a small number of directions concentrate disproportionate energy, while the remaining dimensions form a broad semantic tail. In low-bit training regimes, this geometry becomes numerically unstable. Because blockwise quantization scales are determined by extreme elementwise magnitudes, dominant directions stretch the dynamic range, compressing long-tail semantic variation into narrow numerical bins. We show that this instability is primarily driven by a coherent rank-one mean bias, which constitutes the dominant component of spectral anisotropy in LLM representations. This mean component emerges systematically across layers and training stages and accounts for the majority of extreme activation magnitudes, making it the principal driver of dynamic-range inflation under low precision. Crucially, because the dominant instability is rank-one, it can be eliminated through a simple source-level mean-subtraction operation. This bias-centric conditioning recovers most of the stability benefits of SVD-based spectral methods while requiring only reduction operations and standard quantization kernels. Empirical results on FP4 (W4A4G4) training show that mean removal substantially narrows the loss gap to BF16 and restores downstream performance, providing a hardware-efficient path to stable low-bit LLM training.</p> + https://arxiv.org/abs/2603.10444 + Wed, 11 Mar 2026 05:59:12 +0000 - Code2Math: Can Your Code Agent Effectively Evolve Math Problems Through Exploration? - https://arxiv.org/abs/2603.03202 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03202.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dadi Guo, Yuejin Xie, Qingyu Liu, Jiayu Liu, Zhiyuan Fan, Qihan Ren, Shuai Shao, Tianyi Zhou, Dongrui Liu, Yi R. Fung</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> As large language models (LLMs) advance their mathematical capabilities toward the IMO level, the scarcity of challenging, high-quality problems for training and evaluation has become a significant bottleneck. Simultaneously, recent code agents have demonstrated sophisticated skills in agentic coding and reasoning, suggesting that code execution can serve as a scalable environment for mathematical experimentation. In this paper, we investigate the potential of code agents to autonomously evolve existing math problems into more complex variations. We introduce a multi-agent framework designed to perform problem evolution while validating the solvability and increased difficulty of the generated problems. Our experiments demonstrate that, given sufficient test-time exploration, code agents can synthesize new, solvable problems that are structurally distinct from and more challenging than the originals. This work provides empirical evidence that code-driven agents can serve as a viable mechanism for synthesizing high-difficulty mathematical reasoning problems within scalable computational environments. Our data is available at https://github.com/TarferSoul/Code2Math.</p> - https://arxiv.org/abs/2603.03202 - Tue, 03 Mar 2026 17:55:10 +0000 + Understanding by Reconstruction: Reversing the Software Development Process for LLM Pretraining + https://arxiv.org/abs/2603.11103 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11103.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhiyuan Zeng, Yichi Zhang, Yong Shan, Kai Hua, Siyuan Fang, Zhaiyu Liu, Jiaheng Liu, Haozhe Wang, Yining Zheng, Ming Ding, Ke Shen, Ge Zhang, Wenhao Huang, Xipeng Qiu</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> While Large Language Models (LLMs) have achieved remarkable success in code generation, they often struggle with the deep, long-horizon reasoning required for complex software engineering. We attribute this limitation to the nature of standard pre-training data: static software repositories represent only the terminal state of an intricate intellectual process, abstracting away the intermediate planning, debugging, and iterative refinement. To bridge this gap, we propose a novel paradigm: understanding via reconstruction. We hypothesize that reverse-engineering the latent agentic trajectories -- the planning, reasoning, and debugging steps -- behind static repositories provides a far richer supervision signal than raw code alone. To operationalize this, we introduce a framework that synthesizes these trajectories using a multi-agent simulation. This process is grounded in the structural realities of the source repositories (e.g., dependency graphs and file hierarchies) to ensure fidelity. Furthermore, to guarantee the logical rigor of the synthetic data, we employ a search-based optimization technique that iteratively refines the Chain-of-Thought (CoT) reasoning to maximize the likelihood of the ground-truth code. Empirical results demonstrate that continuous pre-training on these reconstructed trajectories significantly enhances Llama-3-8B's performance across diverse benchmarks, including long-context understanding, coding proficiency, and agentic capabilities.</p> + https://arxiv.org/abs/2603.11103 + Wed, 11 Mar 2026 09:23:20 +0000 - Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use - https://arxiv.org/abs/2603.03205 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03205.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Aradhye Agarwal, Gurdit Siyan, Yash Pandya, Joykirat Singh, Akshay Nambi, Ahmed Awadallah</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Agentic language models operate in a fundamentally different safety regime than chat models: they must plan, call tools, and execute long-horizon actions where a single misstep, such as accessing files or entering credentials, can cause irreversible harm. Existing alignment methods, largely optimized for static generation and task completion, break down in these settings due to sequential decision-making, adversarial tool feedback, and overconfident intermediate reasoning. We introduce MOSAIC, a post-training framework that aligns agents for safe multi-step tool use by making safety decisions explicit and learnable. MOSAIC structures inference as a plan, check, then act or refuse loop, with explicit safety reasoning and refusal as first-class actions. To train without trajectory-level labels, we use preference-based reinforcement learning with pairwise trajectory comparisons, which captures safety distinctions often missed by scalar rewards. We evaluate MOSAIC zero-shot across three model families, Qwen2.5-7B, Qwen3-4B-Thinking, and Phi-4, and across out-of-distribution benchmarks spanning harmful tasks, prompt injection, benign tool use, and cross-domain privacy leakage. MOSAIC reduces harmful behavior by up to 50%, increases harmful-task refusal by over 20% on injection attacks, cuts privacy leakage, and preserves or improves benign task performance, demonstrating robust generalization across models, domains, and agentic settings.</p> - https://arxiv.org/abs/2603.03205 - Tue, 03 Mar 2026 17:59:35 +0000 + Are Video Reasoning Models Ready to Go Outside? + https://arxiv.org/abs/2603.10652 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10652.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yangfan He, Changgyu Boo, Jaehong Yoon</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> In real-world deployment, vision-language models often encounter disturbances such as weather, occlusion, and camera motion. Under such conditions, their understanding and reasoning degrade substantially, revealing a gap between clean, controlled (i.e., unperturbed) evaluation settings and real-world robustness. To address this limitation, we propose ROVA, a novel training framework that improves robustness by modeling a robustness-aware consistency reward under spatio-temporal corruptions. ROVA introduces a difficulty-aware online training strategy that prioritizes informative samples based on the model's evolving capability. Specifically, it continuously re-estimates sample difficulty via self-reflective evaluation, enabling adaptive training with a robustness-aware consistency reward. We also introduce PVRBench, a new benchmark that injects real-world perturbations into embodied video datasets to assess both accuracy and reasoning quality under realistic disturbances. We evaluate ROVA and baselines on PVRBench, UrbanVideo, and VisBench, where open-source and proprietary models suffer up to 35% and 28% drops in accuracy and reasoning under realistic perturbations. ROVA effectively mitigates performance degradation, boosting relative accuracy by at least 24% and reasoning by over 9% compared with baseline models (QWen2.5/3-VL, InternVL2.5, Embodied-R). These gains transfer to clean standard benchmarks, yielding consistent improvements.</p> + https://arxiv.org/abs/2603.10652 + Wed, 11 Mar 2026 11:10:52 +0000 - UniG2U-Bench: Do Unified Models Advance Multimodal Understanding? - https://arxiv.org/abs/2603.03241 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03241.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zimo Wen, Boxiu Li, Wanbo Zhang, Junxiang Lei, Xiaoyu Chen, Yijia Fan, Qi Zhang, Yujiang Wang, Lili Qiu, Bo Li, Ziwei Liu, Caihua Shan, Yifan Yang, Yifei Shen</p><p><b>Upvotes:</b> 79</p><p><b>Summary:</b> Unified multimodal models have recently demonstrated strong generative capabilities, yet whether and when generation improves understanding remains unclear. Existing benchmarks lack a systematic exploration of the specific tasks where generation facilitates understanding. To this end, we introduce UniG2U-Bench, a comprehensive benchmark categorizing generation-to-understanding (G2U) evaluation into 7 regimes and 30 subtasks, requiring varying degrees of implicit or explicit visual transformations. Extensive evaluation of over 30 models reveals three core findings: 1) Unified models generally underperform their base Vision-Language Models (VLMs), and Generate-then-Answer (GtA) inference typically degrades performance relative to direct inference. 2) Consistent enhancements emerge in spatial intelligence, visual illusions, or multi-round reasoning subtasks, where enhanced spatial and shape perception, as well as multi-step intermediate image states, prove beneficial. 3) Tasks with similar reasoning structures and models sharing architectures exhibit correlated behaviors, suggesting that generation-understanding coupling induces class-consistent inductive biases over tasks, pretraining data, and model architectures. These findings highlight the necessity for more diverse training data and novel paradigms to fully unlock the potential of unified multimodal modeling.</p> - https://arxiv.org/abs/2603.03241 - Tue, 03 Mar 2026 18:36:16 +0000 + Prism-Δ: Differential Subspace Steering for Prompt Highlighting in Large Language Models + https://arxiv.org/abs/2603.10705 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10705.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuyao Ge, Shenghua Liu, Yiwei Wang, Tianyu Liu, Baolong Bi, Lingrui Mei, Jiayu Yao, Jiafeng Guo, Xueqi Cheng</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Prompt highlighting steers a large language model to prioritize user-specified text spans during generation. A key challenge is extracting steering directions that capture the difference between relevant and irrelevant contexts, rather than shared structural patterns common to both. We propose PRISM-Δ (Projection-based Relevance-Informed Steering Method), which decomposes the difference between positive and negative cross-covariance matrices to maximize discriminative energy while eliminating shared directions. Each attention head receives a continuous softplus importance weight, letting weak-but-useful heads contribute at reduced strength. The framework extends naturally to Value representations, capturing content-channel signal that Key-only methods leave unused. Across four benchmarks and five models, PRISM-Δ matches or exceeds the best existing method on 19 of 20 configurations, with relative gains up to +10.6%, while halving the fluency cost of steering. PRISM-Δ also scales to long-context retrieval, outperforming the best existing method by up to +4.8% relative gain. PRISM-Δ is compatible with FlashAttention and adds negligible memory overhead.</p> + https://arxiv.org/abs/2603.10705 + Wed, 11 Mar 2026 12:24:45 +0000 - Beyond Language Modeling: An Exploration of Multimodal Pretraining - https://arxiv.org/abs/2603.03276 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03276.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shengbang Tong, David Fan, John Nguyen, Ellis Brown, Gaoyue Zhou, Shengyi Qian, Boyang Zheng, Théophane Vallaeys, Junlin Han, Rob Fergus, Naila Murray, Marjan Ghazvininejad, Mike Lewis, Nicolas Ballas, Amir Bar, Michael Rabbat, Jakob Verbeek, Luke Zettlemoyer, Koustuv Sinha, Yann LeCun, Saining Xie</p><p><b>Upvotes:</b> 73</p><p><b>Summary:</b> The visual world offers a critical axis for advancing foundation models beyond language. Despite growing interest in this direction, the design space for native multimodal models remains opaque. We provide empirical clarity through controlled, from-scratch pretraining experiments, isolating the factors that govern multimodal pretraining without interference from language pretraining. We adopt the Transfusion framework, using next-token prediction for language and diffusion for vision, to train on diverse data including text, video, image-text pairs, and even action-conditioned video. Our experiments yield four key insights: (i) Representation Autoencoder (RAE) provides an optimal unified visual representation by excelling at both visual understanding and generation; (ii) visual and language data are complementary and yield synergy for downstream capabilities; (iii) unified multimodal pretraining leads naturally to world modeling, with capabilities emerging from general training; and (iv) Mixture-of-Experts (MoE) enables efficient and effective multimodal scaling while naturally inducing modality specialization. Through IsoFLOP analysis, we compute scaling laws for both modalities and uncover a scaling asymmetry: vision is significantly more data-hungry than language. We demonstrate that the MoE architecture harmonizes this scaling asymmetry by providing the high model capacity required by language while accommodating the data-intensive nature of vision, paving the way for truly unified multimodal models.</p> - https://arxiv.org/abs/2603.03276 - Tue, 03 Mar 2026 18:58:00 +0000 + Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers + https://arxiv.org/abs/2603.10744 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10744.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wenhao Sun, Ji Li, Zhaoqiang Liu</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Diffusion Transformers have established a new state-of-the-art in image synthesis, but the high computational cost of iterative sampling severely hampers their practical deployment. While existing acceleration methods often focus on the temporal domain, they overlook the substantial spatial redundancy inherent in the generative process, where global structures emerge long before fine-grained details are formed. The uniform computational treatment of all spatial regions represents a critical inefficiency. In this paper, we introduce Just-in-Time (JiT), a novel training-free framework that addresses this challenge by acceleration in the spatial domain. JiT formulates a spatially approximated generative ordinary differential equation (ODE) that drives the full latent state evolution based on computations from a dynamically selected, sparse subset of anchor tokens. To ensure seamless transitions as new tokens are incorporated to expand the dimensions of the latent state, we propose a deterministic micro-flow, a simple and effective finite-time ODE that maintains both structural coherence and statistical correctness. Extensive experiments on the state-of-the-art FLUX.1-dev model demonstrate that JiT achieves up to a 7x speedup with nearly lossless performance, significantly outperforming existing acceleration methods and establishing a new and superior trade-off between inference speed and generation fidelity.</p> + https://arxiv.org/abs/2603.10744 + Wed, 11 Mar 2026 13:16:41 +0000 - CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance - https://arxiv.org/abs/2603.03281 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03281.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hanyang Wang, Yiyang Liu, Jiawei Chi, Fangfu Liu, Ran Xue, Yueqi Duan</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Classifier-Free Guidance (CFG) has emerged as a central approach for enhancing semantic alignment in flow-based diffusion models. In this paper, we explore a unified framework called CFG-Ctrl, which reinterprets CFG as a control applied to the first-order continuous-time generative flow, using the conditional-unconditional discrepancy as an error signal to adjust the velocity field. From this perspective, we summarize vanilla CFG as a proportional controller (P-control) with fixed gain, and typical follow-up variants develop extended control-law designs derived from it. However, existing methods mainly rely on linear control, inherently leading to instability, overshooting, and degraded semantic fidelity especially on large guidance scales. To address this, we introduce Sliding Mode Control CFG (SMC-CFG), which enforces the generative flow toward a rapidly convergent sliding manifold. Specifically, we define an exponential sliding mode surface over the semantic prediction error and introduce a switching control term to establish nonlinear feedback-guided correction. Moreover, we provide a Lyapunov stability analysis to theoretically support finite-time convergence. Experiments across text-to-image generation models including Stable Diffusion 3.5, Flux, and Qwen-Image demonstrate that SMC-CFG outperforms standard CFG in semantic alignment and enhances robustness across a wide range of guidance scales. Project Page: https://hanyang-21.github.io/CFG-Ctrl</p> - https://arxiv.org/abs/2603.03281 - Tue, 03 Mar 2026 18:59:48 +0000 + CodePercept: Code-Grounded Visual STEM Perception for MLLMs + https://arxiv.org/abs/2603.10757 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10757.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tongkun Guan, Zhibo Yang, Jianqiang Wan, Mingkun Yang, Zhengtao Guo, Zijian Hu, Ruilin Luo, Ruize Chen, Songtao Jiang, Peng Wang, Wei Shen, Junyang Lin, Xiaokang Yang</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> When MLLMs fail at Science, Technology, Engineering, and Mathematics (STEM) visual reasoning, a fundamental question arises: is it due to perceptual deficiencies or reasoning limitations? Through systematic scaling analysis that independently scales perception and reasoning components, we uncover a critical insight: scaling perception consistently outperforms scaling reasoning. This reveals perception as the true lever limiting current STEM visual reasoning. Motivated by this insight, our work focuses on systematically enhancing the perception capabilities of MLLMs by establishing code as a powerful perceptual medium--executable code provides precise semantics that naturally align with the structured nature of STEM visuals. Specifically, we construct ICC-1M, a large-scale dataset comprising 1M Image-Caption-Code triplets that materializes this code-as-perception paradigm through two complementary approaches: (1) Code-Grounded Caption Generation treats executable code as ground truth for image captions, eliminating the hallucinations inherent in existing knowledge distillation methods; (2) STEM Image-to-Code Translation prompts models to generate reconstruction code, mitigating the ambiguity of natural language for perception enhancement. To validate this paradigm, we further introduce STEM2Code-Eval, a novel benchmark that directly evaluates visual perception in STEM domains. Unlike existing work relying on problem-solving accuracy as a proxy that only measures problem-relevant understanding, our benchmark requires comprehensive visual comprehension through executable code generation for image reconstruction, providing deterministic and verifiable assessment. Code is available at https://github.com/TongkunGuan/Qwen-CodePercept.</p> + https://arxiv.org/abs/2603.10757 + Wed, 11 Mar 2026 13:32:58 +0000 - Utonia: Toward One Encoder for All Point Clouds - https://arxiv.org/abs/2603.03283 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03283.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yujia Zhang, Xiaoyang Wu, Yunhan Yang, Xianzhe Fan, Han Li, Yuechen Zhang, Zehao Huang, Naiyan Wang, Hengshuang Zhao</p><p><b>Upvotes:</b> 148</p><p><b>Summary:</b> We dream of a future where point clouds from all domains can come together to shape a single model that benefits them all. Toward this goal, we present Utonia, a first step toward training a single self-supervised point transformer encoder across diverse domains, spanning remote sensing, outdoor LiDAR, indoor RGB-D sequences, object-centric CAD models, and point clouds lifted from RGB-only videos. Despite their distinct sensing geometries, densities, and priors, Utonia learns a consistent representation space that transfers across domains. This unification improves perception capability while revealing intriguing emergent behaviors that arise only when domains are trained jointly. Beyond perception, we observe that Utonia representations can also benefit embodied and multimodal reasoning: conditioning vision-language-action policies on Utonia features improves robotic manipulation, and integrating them into vision-language models yields gains on spatial reasoning. We hope Utonia can serve as a step toward foundation models for sparse 3D data, and support downstream applications in AR/VR, robotics, and autonomous driving.</p> - https://arxiv.org/abs/2603.03283 - Tue, 03 Mar 2026 18:59:58 +0000 + V_{0.5}: Generalist Value Model as a Prior for Sparse RL Rollouts + https://arxiv.org/abs/2603.10848 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10848.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yi-Kai Zhang, Yueqing Sun, Hongyan Hao, Qi Gu, Xunliang Cai, De-Chuan Zhan, Han-Jia Ye</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> In Reinforcement Learning with Verifiable Rewards (RLVR), constructing a robust advantage baseline is critical for policy gradients, effectively guiding the policy model to reinforce desired behaviors. Recent research has introduced Generalist Value Models (such as V_0), which achieve pre-trained value estimation by explicitly encoding model capabilities in-context, eliminating the need to synchronously update the value model alongside the policy model. In this paper, we propose V_{0.5}, which adaptively fuses the baseline predicted by such value model (acting as a prior) with the empirical mean derived from sparse rollouts. This constructs a robust baseline that balances computational efficiency with extremely low variance. Specifically, we introduce a real-time statistical testing and dynamic budget allocation. This balances the high variance caused by sparse sampling against the systematic bias (or hallucinations) inherent in the value model's prior. By constructing a hypothesis test to evaluate the prior's reliability in real-time, the system dynamically allocates additional rollout budget on demand. This mechanism minimizes the baseline estimator's Mean Squared Error (MSE), guaranteeing stable policy gradients, even under extreme sparsity with a group size of 4. Extensive evaluations across six mathematical reasoning benchmarks demonstrate that V_{0.5} significantly outperforms GRPO and DAPO, achieving faster convergence and over some 10% performance improvement.</p> + https://arxiv.org/abs/2603.10848 + Wed, 11 Mar 2026 14:57:41 +0000 - Proact-VL: A Proactive VideoLLM for Real-Time AI Companions - https://arxiv.org/abs/2603.03447 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03447.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Weicai Yan, Yuhong Dai, Qi Ran, Haodong Li, Wang Lin, Hao Liao, Xing Xie, Tao Jin, Jianxun Lian</p><p><b>Upvotes:</b> 25</p><p><b>Summary:</b> Proactive and real-time interactive experiences are essential for human-like AI companions, yet face three key challenges: (1) achieving low-latency inference under continuous streaming inputs, (2) autonomously deciding when to respond, and (3) controlling both quality and quantity of generated content to meet real-time constraints. In this work, we instantiate AI companions through two gaming scenarios, commentator and guide, selected for their suitability for automatic evaluation. We introduce the Live Gaming Benchmark, a large-scale dataset with three representative scenarios: solo commentary, co-commentary, and user guidance, and present Proact-VL, a general framework that shapes multimodal language models into proactive, real-time interactive agents capable of human-like environment perception and interaction. Extensive experiments show Proact-VL achieves superior response latency and quality while maintaining strong video understanding capabilities, demonstrating its practicality for real-time interactive applications.</p> - https://arxiv.org/abs/2603.03447 - Tue, 03 Mar 2026 19:02:46 +0000 + LLM2Vec-Gen: Generative Embeddings from Large Language Models + https://arxiv.org/abs/2603.10913 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.10913.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Parishad BehnamGhader, Vaibhav Adlakha, Fabian David Schmidt, Nicolas Chapados, Marius Mosbach, Siva Reddy</p><p><b>Upvotes:</b> 34</p><p><b>Summary:</b> LLM-based text embedders typically encode the semantic content of their input. However, embedding tasks require mapping diverse inputs to similar outputs. Typically, this input-output is addressed by training embedding models with paired data using contrastive learning. In this work, we propose a novel self-supervised approach, LLM2Vec-Gen, which adopts a different paradigm: rather than encoding the input, we learn to represent the model's potential response. Specifically, we add trainable special tokens to the LLM's vocabulary, append them to input, and optimize them to represent the LLM's response in a fixed-length sequence. Training is guided by the LLM's own completion for the query, along with an unsupervised embedding teacher that provides distillation targets. This formulation helps to bridge the input-output gap and transfers LLM capabilities such as safety alignment and reasoning to embedding tasks. Crucially, the LLM backbone remains frozen and training requires only unlabeled queries. LLM2Vec-Gen achieves state-of-the-art self-supervised performance on the Massive Text Embedding Benchmark (MTEB), improving by 9.3% over the best unsupervised embedding teacher. We also observe up to 43.2% reduction in harmful content retrieval and 29.3% improvement in reasoning capabilities for embedding tasks. Finally, the learned embeddings are interpretable and can be decoded into text to reveal their semantic content.</p> + https://arxiv.org/abs/2603.10913 + Wed, 11 Mar 2026 15:58:47 +0000 - InfinityStory: Unlimited Video Generation with World Consistency and Character-Aware Shot Transitions - https://arxiv.org/abs/2603.03646 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03646.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Mohamed Elmoghany, Liangbing Zhao, Xiaoqian Shen, Subhojyoti Mukherjee, Yang Zhou, Gang Wu, Viet Dac Lai, Seunghyun Yoon, Ryan Rossi, Abdullah Rashwan, Puneet Mathur, Varun Manjunatha, Daksh Dangi, Chien Nguyen, Nedim Lipka, Trung Bui, Krishna Kumar Singh, Ruiyi Zhang, Xiaolei Huang, Jaemin Cho, Yu Wang, Namyong Park, Zhengzhong Tu, Hongjie Chen, Hoda Eldardiry, Nesreen Ahmed, Thien Nguyen, Dinesh Manocha, Mohamed Elhoseiny, Franck Dernoncourt</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Generating long-form storytelling videos with consistent visual narratives remains a significant challenge in video synthesis. We present a novel framework, dataset, and a model that address three critical limitations: background consistency across shots, seamless multi-subject shot-to-shot transitions, and scalability to hour-long narratives. Our approach introduces a background-consistent generation pipeline that maintains visual coherence across scenes while preserving character identity and spatial relationships. We further propose a transition-aware video synthesis module that generates smooth shot transitions for complex scenarios involving multiple subjects entering or exiting frames, going beyond the single-subject limitations of prior work. To support this, we contribute with a synthetic dataset of 10,000 multi-subject transition sequences covering underrepresented dynamic scene compositions. On VBench, InfinityStory achieves the highest Background Consistency (88.94), highest Subject Consistency (82.11), and the best overall average rank (2.80), showing improved stability, smoother transitions, and better temporal coherence.</p> - https://arxiv.org/abs/2603.03646 - Wed, 04 Mar 2026 02:10:32 +0000 + Meta-Reinforcement Learning with Self-Reflection for Agentic Search + https://arxiv.org/abs/2603.11327 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11327.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Teng Xiao, Yige Yuan, Hamish Ivison, Huaisheng Zhu, Faeze Brahman, Nathan Lambert, Pradeep Dasigi, Noah A. Smith, Hannaneh Hajishirzi</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> This paper introduces MR-Search, an in-context meta reinforcement learning (RL) formulation for agentic search with self-reflection. Instead of optimizing a policy within a single independent episode with sparse rewards, MR-Search trains a policy that conditions on past episodes and adapts its search strategy across episodes. MR-Search learns to learn a search strategy with self-reflection, allowing search agents to improve in-context exploration at test-time. Specifically, MR-Search performs cross-episode exploration by generating explicit self-reflections after each episode and leveraging them as additional context to guide subsequent attempts, thereby promoting more effective exploration during test-time. We further introduce a multi-turn RL algorithm that estimates a dense relative advantage at the turn level, enabling fine-grained credit assignment on each episode. Empirical results across various benchmarks demonstrate the advantages of MR-Search over baselines based RL, showing strong generalization and relative improvements of 9.2% to 19.3% across eight benchmarks. Our code and data are available at https://github.com/tengxiao1/MR-Search.</p> + https://arxiv.org/abs/2603.11327 + Wed, 11 Mar 2026 21:40:26 +0000 - MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier - https://arxiv.org/abs/2603.03756 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03756.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zonglin Yang, Lidong Bing</p><p><b>Upvotes:</b> 82</p><p><b>Summary:</b> While large language models (LLMs) show promise in scientific discovery, existing research focuses on inference or feedback-driven training, leaving the direct modeling of the generative reasoning process, P(hypothesis|background) (P(h|b)), unexplored. We demonstrate that directly training P(h|b) is mathematically intractable due to the combinatorial complexity (O(N^k)) inherent in retrieving and composing inspirations from a vast knowledge base. To break this barrier, we introduce MOOSE-Star, a unified framework enabling tractable training and scalable inference. In the best case, MOOSE-Star reduces complexity from exponential to logarithmic (O(log N)) by (1) training on decomposed subtasks derived from the probabilistic equation of discovery, (2) employing motivation-guided hierarchical search to enable logarithmic retrieval and prune irrelevant subspaces, and (3) utilizing bounded composition for robustness against retrieval noise. To facilitate this, we release TOMATO-Star, a dataset of 108,717 decomposed papers (38,400 GPU hours) for training. Furthermore, we show that while brute-force sampling hits a ''complexity wall,'' MOOSE-Star exhibits continuous test-time scaling.</p> - https://arxiv.org/abs/2603.03756 - Wed, 04 Mar 2026 06:11:18 +0000 + ShotVerse: Advancing Cinematic Camera Control for Text-Driven Multi-Shot Video Creation + https://arxiv.org/abs/2603.11421 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11421.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Songlin Yang, Zhe Wang, Xuyi Yang, Songchun Zhang, Xianghao Kong, Taiyi Wu, Xiaotong Zhao, Ran Zhang, Alan Zhao, Anyi Rao</p><p><b>Upvotes:</b> 28</p><p><b>Summary:</b> Text-driven video generation has democratized film creation, but camera control in cinematic multi-shot scenarios remains a significant block. Implicit textual prompts lack precision, while explicit trajectory conditioning imposes prohibitive manual overhead and often triggers execution failures in current models. To overcome this bottleneck, we propose a data-centric paradigm shift, positing that aligned (Caption, Trajectory, Video) triplets form an inherent joint distribution that can connect automated plotting and precise execution. Guided by this insight, we present ShotVerse, a "Plan-then-Control" framework that decouples generation into two collaborative agents: a VLM (Vision-Language Model)-based Planner that leverages spatial priors to obtain cinematic, globally aligned trajectories from text, and a Controller that renders these trajectories into multi-shot video content via a camera adapter. Central to our approach is the construction of a data foundation: we design an automated multi-shot camera calibration pipeline aligns disjoint single-shot trajectories into a unified global coordinate system. This facilitates the curation of ShotVerse-Bench, a high-fidelity cinematic dataset with a three-track evaluation protocol that serves as the bedrock for our framework. Extensive experiments demonstrate that ShotVerse effectively bridges the gap between unreliable textual control and labor-intensive manual plotting, achieving superior cinematic aesthetics and generating multi-shot videos that are both camera-accurate and cross-shot consistent.</p> + https://arxiv.org/abs/2603.11421 + Thu, 12 Mar 2026 01:27:08 +0000 - T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning - https://arxiv.org/abs/2603.03790 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03790.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qinsi Wang, Hancheng Ye, Jinhee Kim, Jinghan Ke, Yifei Wang, Martin Kuo, Zishan Shao, Dongting Li, Yueqian Lin, Ting Jiang, Chiyue Wei, Qi Qian, Wei Wen, Helen Li, Yiran Chen</p><p><b>Upvotes:</b> 109</p><p><b>Summary:</b> Think about how human handles complex reading tasks: marking key points, inferring their relationships, and structuring information to guide understanding and responses. Likewise, can a large language model benefit from text structure to enhance text-processing performance? To explore it, in this work, we first introduce Structure of Thought (SoT), a prompting technique that explicitly guides models to construct intermediate text structures, consistently boosting performance across eight tasks and three model families. Building upon this insight, we present T2S-Bench, the first benchmark designed to evaluate and improve text-to-structure capabilities of models. T2S-Bench includes 1.8K samples across 6 scientific domains and 32 structural types, rigorously constructed to ensure accuracy, fairness, and quality. Evaluation on 45 mainstream models reveals substantial improvement potential: the average accuracy on the multi-hop reasoning task is only 52.1%, and even the most advanced model achieves 58.1% node accuracy in end-to-end extraction. Furthermore, on Qwen2.5-7B-Instruct, SoT alone yields an average +5.7% improvement across eight diverse text-processing tasks, and fine-tuning on T2S-Bench further increases this gain to +8.6%. These results highlight the value of explicit text structuring and the complementary contributions of SoT and T2S-Bench. Dataset and eval code have been released at https://t2s-bench.github.io/T2S-Bench-Page/.</p> - https://arxiv.org/abs/2603.03790 - Wed, 04 Mar 2026 07:05:09 +0000 + Tiny Aya: Bridging Scale and Multilingual Depth + https://arxiv.org/abs/2603.11510 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11510.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Alejandro R. Salamanca, Diana Abagyan, Daniel D'souza, Ammar Khairi, David Mora, Saurabh Dash, Viraat Aryabumi, Sara Rajaee, Mehrnaz Mofakhami, Ananya Sahu, Thomas Euyang, Brittawnya Prince, Madeline Smith, Hangyu Lin, Acyr Locatelli, Sara Hooker, Tom Kocmi, Aidan Gomez, Ivan Zhang, Phil Blunsom, Nick Frosst, Joelle Pineau, Beyza Ermis, Ahmet Üstün, Julia Kreutzer, Marzieh Fadaee</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Tiny Aya redefines what a small multilingual language model can achieve. Trained on 70 languages and refined through region-aware posttraining, it delivers state-of-the-art in translation quality, strong multilingual understanding, and high-quality target-language generation, all with just 3.35B parameters. The release includes a pretrained foundation model, a globally balanced instruction-tuned variant, and three region-specialized models targeting languages from Africa, South Asia, Europe, Asia-Pacific, and West Asia. This report details the training strategy, data composition, and comprehensive evaluation framework behind Tiny Aya, and presents an alternative scaling path for multilingual AI: one centered on efficiency, balanced performance across languages, and practical deployment.</p> + https://arxiv.org/abs/2603.11510 + Thu, 12 Mar 2026 03:53:13 +0000 - Phi-4-reasoning-vision-15B Technical Report - https://arxiv.org/abs/2603.03975 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03975.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jyoti Aneja, Michael Harrison, Neel Joshi, Tyler LaBonte, John Langford, Eduardo Salinas</p><p><b>Upvotes:</b> 19</p><p><b>Summary:</b> We present Phi-4-reasoning-vision-15B, a compact open-weight multimodal reasoning model, and share the motivations, design choices, experiments, and learnings that informed its development. Our goal is to contribute practical insight to the research community on building smaller, efficient multimodal reasoning models and to share the result of these learnings as an open-weight model that is good at common vision and language tasks and excels at scientific and mathematical reasoning and understanding user interfaces. Our contributions include demonstrating that careful architecture choices and rigorous data curation enable smaller, open-weight multimodal models to achieve competitive performance with significantly less training and inference-time compute and tokens. The most substantial improvements come from systematic filtering, error correction, and synthetic augmentation -- reinforcing that data quality remains the primary lever for model performance. Systematic ablations show that high-resolution, dynamic-resolution encoders yield consistent improvements, as accurate perception is a prerequisite for high-quality reasoning. Finally, a hybrid mix of reasoning and non-reasoning data with explicit mode tokens allows a single model to deliver fast direct answers for simpler tasks and chain-of-thought reasoning for complex problems.</p> - https://arxiv.org/abs/2603.03975 - Wed, 04 Mar 2026 12:16:53 +0000 + Mobile-GS: Real-time Gaussian Splatting for Mobile Devices + https://arxiv.org/abs/2603.11531 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11531.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiaobiao Du, Yida Wang, Kun Zhan, Xin Yu</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> 3D Gaussian Splatting (3DGS) has emerged as a powerful representation for high-quality rendering across a wide range of applications.However, its high computational demands and large storage costs pose significant challenges for deployment on mobile devices. In this work, we propose a mobile-tailored real-time Gaussian Splatting method, dubbed Mobile-GS, enabling efficient inference of Gaussian Splatting on edge devices. Specifically, we first identify alpha blending as the primary computational bottleneck, since it relies on the time-consuming Gaussian depth sorting process. To solve this issue, we propose a depth-aware order-independent rendering scheme that eliminates the need for sorting, thereby substantially accelerating rendering. Although this order-independent rendering improves rendering speed, it may introduce transparency artifacts in regions with overlapping geometry due to the scarcity of rendering order. To address this problem, we propose a neural view-dependent enhancement strategy, enabling more accurate modeling of view-dependent effects conditioned on viewing direction, 3D Gaussian geometry, and appearance attributes. In this way, Mobile-GS can achieve both high-quality and real-time rendering. Furthermore, to facilitate deployment on memory-constrained mobile platforms, we also introduce first-order spherical harmonics distillation, a neural vector quantization technique, and a contribution-based pruning strategy to reduce the number of Gaussian primitives and compress the 3D Gaussian representation with the assistance of neural networks. Extensive experiments demonstrate that our proposed Mobile-GS achieves real-time rendering and compact model size while preserving high visual quality, making it well-suited for mobile applications.</p> + https://arxiv.org/abs/2603.11531 + Thu, 12 Mar 2026 04:33:04 +0000 - RIVER: A Real-Time Interaction Benchmark for Video LLMs - https://arxiv.org/abs/2603.03985 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.03985.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yansong Shi, Qingsong Zhao, Tianxiang Jiang, Xiangyu Zeng, Yi Wang, Limin Wang</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> The rapid advancement of multimodal large language models has demonstrated impressive capabilities, yet nearly all operate in an offline paradigm, hindering real-time interactivity. Addressing this gap, we introduce the Real-tIme Video intERaction Bench (RIVER Bench), designed for evaluating online video comprehension. RIVER Bench introduces a novel framework comprising Retrospective Memory, Live-Perception, and Proactive Anticipation tasks, closely mimicking interactive dialogues rather than responding to entire videos at once. We conducted detailed annotations using videos from diverse sources and varying lengths, and precisely defined the real-time interactive format. Evaluations across various model categories reveal that while offline models perform well in single question-answering tasks, they struggle with real-time processing. Addressing the limitations of existing models in online video interaction, especially their deficiencies in long-term memory and future perception, we proposed a general improvement method that enables models to interact with users more flexibly in real time. We believe this work will significantly advance the development of real-time interactive video understanding models and inspire future research in this emerging field. Datasets and code are publicly available at https://github.com/OpenGVLab/RIVER.</p> - https://arxiv.org/abs/2603.03985 - Wed, 04 Mar 2026 12:29:34 +0000 + WeEdit: A Dataset, Benchmark and Glyph-Guided Framework for Text-centric Image Editing + https://arxiv.org/abs/2603.11593 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.11593.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hui Zhang, Juntao Liu, Zongkai Liu, Liqiang Niu, Fandong Meng, Zuxuan Wu, Yu-Gang Jiang</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> Instruction-based image editing aims to modify specific content within existing images according to user-provided instructions while preserving non-target regions. Beyond traditional object- and style-centric manipulation, text-centric image editing focuses on modifying, translating, or rearranging textual elements embedded within images. However, existing leading models often struggle to execute complex text editing precisely, frequently producing blurry or hallucinated characters. We attribute these failures primarily to the lack of specialized training paradigms tailored for text-centric editing, as well as the absence of large-scale datasets and standardized benchmarks necessary for a closed-loop training and evaluation system. To address these limitations, we present WeEdit, a systematic solution encompassing a scalable data construction pipeline, two benchmarks, and a tailored two-stage training strategy. Specifically, we propose a novel HTML-based automatic editing pipeline, which generates 330K training pairs covering diverse editing operations and 15 languages, accompanied by standardized bilingual and multilingual benchmarks for comprehensive evaluation. On the algorithmic side, we employ glyph-guided supervised fine-tuning to inject explicit spatial and content priors, followed by a multi-objective reinforcement learning stage to align generation with instruction adherence, text clarity, and background preservation. Extensive experiments demonstrate that WeEdit outperforms previous open-source models by a clear margin across diverse editing operations.</p> + https://arxiv.org/abs/2603.11593 + Thu, 12 Mar 2026 06:25:09 +0000 - Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory - https://arxiv.org/abs/2603.04257 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04257.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhenting Wang, Huancheng Chen, Jiayun Wang, Wei Wei</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Large language model (LLM) agents are fundamentally bottlenecked by finite context windows on long-horizon tasks. As trajectories grow, retaining tool outputs and intermediate reasoning in-context quickly becomes infeasible: the working context becomes prohibitively long, eventually exceeds the context budget, and makes distant evidence harder to use even when it is still present. Existing solutions typically shorten context through truncation or running summaries, but these methods are fundamentally lossy because they compress or discard past evidence itself. We introduce Memex, an indexed experience memory mechanism that instead compresses context without discarding evidence. Memex maintains a compact working context consisting of concise structured summaries and stable indices, while storing full-fidelity underlying interactions in an external experience database under those indices. The agent can then decide when to dereference an index and recover the exact past evidence needed for the current subgoal. We optimize both write and read behaviors with our reinforcement learning framework MemexRL, using reward shaping tailored to indexed memory usage under a context budget, so the agent learns what to summarize, what to archive, how to index it, and when to retrieve it. This yields a substantially less lossy form of long-horizon memory than summary-only approaches. We further provide a theoretical analysis showing the potential of the Memex loop to preserve decision quality with bounded dereferencing while keeping effective in-context computation bounded as history grows. Empirically, on challenging long-horizon tasks, Memex agent trained with MemexRL improves task success while using a significantly smaller working context.</p> - https://arxiv.org/abs/2603.04257 - Wed, 04 Mar 2026 16:45:06 +0000 + XSkill: Continual Learning from Experience and Skills in Multimodal Agents + https://arxiv.org/abs/2603.12056 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12056.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Guanyu Jiang, Zhaochen Su, Xiaoye Qu, Yi R., Fung</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> Multimodal agents can now tackle complex reasoning tasks with diverse tools, yet they still suffer from inefficient tool use and inflexible orchestration in open-ended settings. A central challenge is enabling such agents to continually improve without parameter updates by learning from past trajectories. We identify two complementary forms of reusable knowledge essential for this goal: experiences, providing concise action-level guidance for tool selection and decision making, and skills, providing structured task-level guidance for planning and tool use. To this end, we propose XSkill, a dual-stream framework for continual learning from experience and skills in multimodal agents. XSkill grounds both knowledge extraction and retrieval in visual observations. During accumulation, XSkill distills and consolidates experiences and skills from multi-path rollouts via visually grounded summarization and cross-rollout critique. During inference, it retrieves and adapts this knowledge to the current visual context and feeds usage history back into accumulation to form a continual learning loop. Evaluated on five benchmarks across diverse domains with four backbone models, XSkill consistently and substantially outperforms both tool-only and learning-based baselines. Further analysis reveals that the two knowledge streams play complementary roles in influencing the reasoning behaviors of agents and show superior zero-shot generalization.</p> + https://arxiv.org/abs/2603.12056 + Thu, 12 Mar 2026 15:25:57 +0000 - CubeComposer: Spatio-Temporal Autoregressive 4K 360° Video Generation from Perspective Video - https://arxiv.org/abs/2603.04291 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04291.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lingen Li, Guangzhi Wang, Xiaoyu Li, Zhaoyang Zhang, Qi Dou, Jinwei Gu, Tianfan Xue, Ying Shan</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Generating high-quality 360° panoramic videos from perspective input is one of the crucial applications for virtual reality (VR), whereby high-resolution videos are especially important for immersive experience. Existing methods are constrained by computational limitations of vanilla diffusion models, only supporting leq 1K resolution native generation and relying on suboptimal post super-resolution to increase resolution. We introduce CubeComposer, a novel spatio-temporal autoregressive diffusion model that natively generates 4K-resolution 360° videos. By decomposing videos into cubemap representations with six faces, CubeComposer autoregressively synthesizes content in a well-planned spatio-temporal order, reducing memory demands while enabling high-resolution output. Specifically, to address challenges in multi-dimensional autoregression, we propose: (1) a spatio-temporal autoregressive strategy that orchestrates 360° video generation across cube faces and time windows for coherent synthesis; (2) a cube face context management mechanism, equipped with a sparse context attention design to improve efficiency; and (3) continuity-aware techniques, including cube-aware positional encoding, padding, and blending to eliminate boundary seams. Extensive experiments on benchmark datasets demonstrate that CubeComposer outperforms state-of-the-art methods in native resolution and visual quality, supporting practical VR application scenarios. Project page: https://lg-li.github.io/project/cubecomposer</p> - https://arxiv.org/abs/2603.04291 - Wed, 04 Mar 2026 17:06:56 +0000 + Coarse-Guided Visual Generation via Weighted h-Transform Sampling + https://arxiv.org/abs/2603.12057 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12057.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yanghao Wang, Ziqi Jiang, Zhen Wang, Long Chen</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Coarse-guided visual generation, which synthesizes fine visual samples from degraded or low-fidelity coarse references, is essential for various real-world applications. While training-based approaches are effective, they are inherently limited by high training costs and restricted generalization due to paired data collection. Accordingly, recent training-free works propose to leverage pretrained diffusion models and incorporate guidance during the sampling process. However, these training-free methods either require knowing the forward (fine-to-coarse) transformation operator, e.g., bicubic downsampling, or are difficult to balance between guidance and synthetic quality. To address these challenges, we propose a novel guided method by using the h-transform, a tool that can constrain stochastic processes (e.g., sampling process) under desired conditions. Specifically, we modify the transition probability at each sampling timestep by adding to the original differential equation with a drift function, which approximately steers the generation toward the ideal fine sample. To address unavoidable approximation errors, we introduce a noise-level-aware schedule that gradually de-weights the term as the error increases, ensuring both guidance adherence and high-quality synthesis. Extensive experiments across diverse image and video generation tasks demonstrate the effectiveness and generalization of our method.</p> + https://arxiv.org/abs/2603.12057 + Thu, 12 Mar 2026 15:26:19 +0000 - V_1: Unifying Generation and Self-Verification for Parallel Reasoners - https://arxiv.org/abs/2603.04304 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04304.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Harman Singh, Xiuyu Li, Kusha Sareen, Monishwaran Maheswaran, Sijun Tan, Xiaoxia Wu, Junxiong Wang, Alpay Ariyak, Qingyang Wu, Samir Khaki, Rishabh Tiwari, Long Lian, Yucheng Lu, Boyi Li, Alane Suhr, Ben Athiwaratkun, Kurt Keutzer</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Test-time scaling for complex reasoning tasks shows that leveraging inference-time compute, by methods such as independently sampling and aggregating multiple solutions, results in significantly better task outcomes. However, a critical bottleneck is verification: sampling is only effective if correct solutions can be reliably identified among candidates. While existing approaches typically evaluate candidates independently via scalar scoring, we demonstrate that models are substantially stronger at pairwise self-verification. Leveraging this insight, we introduce V_1, a framework that unifies generation and verification through efficient pairwise ranking. V_1 comprises two components: V_1-Infer, an uncertainty-guided algorithm using a tournament-based ranking that dynamically allocates self-verification compute to candidate pairs whose relative correctness is most uncertain; and V_1-PairRL, an RL framework that jointly trains a single model as both generator and pairwise self-verifier, ensuring the verifier adapts to the generator's evolving distribution. On code generation (LiveCodeBench, CodeContests, SWE-Bench) and math reasoning (AIME, HMMT) benchmarks, V_1-Infer improves Pass@1 by up to 10% over pointwise verification and outperforms recent test-time scaling methods while being significantly more efficient. Furthermore, V_1-PairRL achieves 7--9% test-time scaling gains over standard RL and pointwise joint training, and improves base Pass@1 by up to 8.7% over standard RL in a code-generation setting.</p> - https://arxiv.org/abs/2603.04304 - Wed, 04 Mar 2026 17:22:16 +0000 + Automatic Generation of High-Performance RL Environments + https://arxiv.org/abs/2603.12145 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12145.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Seth Karten, Rahul Dev Appapogu, Chi Jin</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Translating complex reinforcement learning (RL) environments into high-performance implementations has traditionally required months of specialized engineering. We present a reusable recipe - a generic prompt template, hierarchical verification, and iterative agent-assisted repair - that produces semantically equivalent high-performance environments for <$10 in compute cost. We demonstrate three distinct workflows across five environments. Direct translation (no prior performance implementation exists): EmuRust (1.5x PPO speedup via Rust parallelism for a Game Boy emulator) and PokeJAX, the first GPU-parallel Pokemon battle simulator (500M SPS random action, 15.2M SPS PPO; 22,320x over the TypeScript reference). Translation verified against existing performance implementations: throughput parity with MJX (1.04x) and 5x over Brax at matched GPU batch sizes (HalfCheetah JAX); 42x PPO (Puffer Pong). New environment creation: TCGJax, the first deployable JAX Pokemon TCG engine (717K SPS random action, 153K SPS PPO; 6.6x over the Python reference), synthesized from a web-extracted specification. At 200M parameters, the environment overhead drops below 4% of training time. Hierarchical verification (property, interaction, and rollout tests) confirms semantic equivalence for all five environments; cross-backend policy transfer confirms zero sim-to-sim gap for all five environments. TCGJax, synthesized from a private reference absent from public repositories, serves as a contamination control for agent pretraining data concerns. The paper contains sufficient detail - including representative prompts, verification methodology, and complete results - that a coding agent could reproduce the translations directly from the manuscript.</p> + https://arxiv.org/abs/2603.12145 + Thu, 12 Mar 2026 16:45:47 +0000 - ArtHOI: Articulated Human-Object Interaction Synthesis by 4D Reconstruction from Video Priors - https://arxiv.org/abs/2603.04338 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04338.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zihao Huang, Tianqi Liu, Zhaoxi Chen, Shaocong Xu, Saining Zhang, Lixing Xiao, Zhiguo Cao, Wei Li, Hao Zhao, Ziwei Liu</p><p><b>Upvotes:</b> 20</p><p><b>Summary:</b> Synthesizing physically plausible articulated human-object interactions (HOI) without 3D/4D supervision remains a fundamental challenge. While recent zero-shot approaches leverage video diffusion models to synthesize human-object interactions, they are largely confined to rigid-object manipulation and lack explicit 4D geometric reasoning. To bridge this gap, we formulate articulated HOI synthesis as a 4D reconstruction problem from monocular video priors: given only a video generated by a diffusion model, we reconstruct a full 4D articulated scene without any 3D supervision. This reconstruction-based approach treats the generated 2D video as supervision for an inverse rendering problem, recovering geometrically consistent and physically plausible 4D scenes that naturally respect contact, articulation, and temporal coherence. We introduce ArtHOI, the first zero-shot framework for articulated human-object interaction synthesis via 4D reconstruction from video priors. Our key designs are: 1) Flow-based part segmentation: leveraging optical flow as a geometric cue to disentangle dynamic from static regions in monocular video; 2) Decoupled reconstruction pipeline: joint optimization of human motion and object articulation is unstable under monocular ambiguity, so we first recover object articulation, then synthesize human motion conditioned on the reconstructed object states. ArtHOI bridges video-based generation and geometry-aware reconstruction, producing interactions that are both semantically aligned and physically grounded. Across diverse articulated scenes (e.g., opening fridges, cabinets, microwaves), ArtHOI significantly outperforms prior methods in contact accuracy, penetration reduction, and articulation fidelity, extending zero-shot interaction synthesis beyond rigid manipulation through reconstruction-informed synthesis.</p> - https://arxiv.org/abs/2603.04338 - Wed, 04 Mar 2026 17:58:04 +0000 + Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections + https://arxiv.org/abs/2603.12180 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12180.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Łukasz Borchmann, Jordy Van Landeghem, Michał Turski, Shreyansh Padarha, Ryan Othniel Kearns, Adam Mahdi, Niels Rogge, Clémentine Fourrier, Siwei Han, Huaxiu Yao, Artemis Llabrés, Yiming Xu, Dimosthenis Karatzas, Hao Zhang, Anupam Datta</p><p><b>Upvotes:</b> 51</p><p><b>Summary:</b> Multimodal agents offer a promising path to automating complex document-intensive workflows. Yet, a critical question remains: do these agents demonstrate genuine strategic reasoning, or merely stochastic trial-and-error search? To address this, we introduce MADQA, a benchmark of 2,250 human-authored questions grounded in 800 heterogeneous PDF documents. Guided by Classical Test Theory, we design it to maximize discriminative power across varying levels of agentic abilities. To evaluate agentic behaviour, we introduce a novel evaluation protocol measuring the accuracy-effort trade-off. Using this framework, we show that while the best agents can match human searchers in raw accuracy, they succeed on largely different questions and rely on brute-force search to compensate for weak strategic planning. They fail to close the nearly 20% gap to oracle performance, persisting in unproductive loops. We release the dataset and evaluation harness to help facilitate the transition from brute-force retrieval to calibrated, efficient reasoning.</p> + https://arxiv.org/abs/2603.12180 + Thu, 12 Mar 2026 17:11:22 +0000 - Helios: Real Real-Time Long Video Generation Model - https://arxiv.org/abs/2603.04379 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04379.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, Xiao Yang, Li Yuan</p><p><b>Upvotes:</b> 136</p><p><b>Summary:</b> We introduce Helios, the first 14B video generation model that runs at 19.5 FPS on a single NVIDIA H100 GPU and supports minute-scale generation while matching the quality of a strong baseline. We make breakthroughs along three key dimensions: (1) robustness to long-video drifting without commonly used anti-drifting heuristics such as self-forcing, error-banks, or keyframe sampling; (2) real-time generation without standard acceleration techniques such as KV-cache, sparse/linear attention, or quantization; and (3) training without parallelism or sharding frameworks, enabling image-diffusion-scale batch sizes while fitting up to four 14B models within 80 GB of GPU memory. Specifically, Helios is a 14B autoregressive diffusion model with a unified input representation that natively supports T2V, I2V, and V2V tasks. To mitigate drifting in long-video generation, we characterize typical failure modes and propose simple yet effective training strategies that explicitly simulate drifting during training, while eliminating repetitive motion at its source. For efficiency, we heavily compress the historical and noisy context and reduce the number of sampling steps, yielding computational costs comparable to -- or lower than -- those of 1.3B video generative models. Moreover, we introduce infrastructure-level optimizations that accelerate both inference and training while reducing memory consumption. Extensive experiments demonstrate that Helios consistently outperforms prior methods on both short- and long-video generation. We plan to release the code, base model, and distilled model to support further development by the community.</p> - https://arxiv.org/abs/2603.04379 - Wed, 04 Mar 2026 18:45:21 +0000 + IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse + https://arxiv.org/abs/2603.12201 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12201.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yushi Bai, Qian Dong, Ting Jiang, Xin Lv, Zhengxiao Du, Aohan Zeng, Jie Tang, Juanzi Li</p><p><b>Upvotes:</b> 39</p><p><b>Summary:</b> Long-context agentic workflows have emerged as a defining use case for large language models, making attention efficiency critical for both inference speed and serving cost. Sparse attention addresses this challenge effectively, and DeepSeek Sparse Attention (DSA) is a representative production-grade solution: a lightweight lightning indexer selects the top-k most relevant tokens per query, reducing core attention from O(L^2) to O(Lk). However, the indexer itself retains O(L^2) complexity and must run independently at every layer, despite the fact that the resulting top-k selections are highly similar across consecutive layers. We present IndexCache, which exploits this cross-layer redundancy by partitioning layers into a small set of Full layers that run their own indexers and a majority of Shared layers that simply reuse the nearest Full layer's top-k indices. We propose two complementary approaches to determine and optimize this configuration. Training-free IndexCache applies a greedy search algorithm that selects which layers to retain indexers by directly minimizing language modeling loss on a calibration set, requiring no weight updates. Training-aware IndexCache introduces a multi-layer distillation loss that trains each retained indexer against the averaged attention distributions of all layers it serves, enabling even simple interleaved patterns to match full-indexer accuracy. Experimental results on a 30B DSA model show that IndexCache can remove 75% of indexer computations with negligible quality degradation, achieving up to 1.82times prefill speedup and 1.48times decode speedup compared to standard DSA. These positive results are further confirmed by our preliminary experiments on the production-scale GLM-5 model (Figure 1).</p> + https://arxiv.org/abs/2603.12201 + Thu, 12 Mar 2026 17:27:21 +0000 - Interactive Benchmarks - https://arxiv.org/abs/2603.04737 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04737.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Baoqing Yue, Zihan Zhu, Yifan Zhang, Jichen Feng, Hufei Yang, Mengdi Wang</p><p><b>Upvotes:</b> 17</p><p><b>Summary:</b> Standard benchmarks have become increasingly unreliable due to saturation, subjectivity, and poor generalization. We argue that evaluating model's ability to acquire information actively is important to assess model's intelligence. We propose Interactive Benchmarks, a unified evaluation paradigm that assesses model's reasoning ability in an interactive process under budget constraints. We instantiate this framework across two settings: Interactive Proofs, where models interact with a judge to deduce objective truths or answers in logic and mathematics; and Interactive Games, where models reason strategically to maximize long-horizon utilities. Our results show that interactive benchmarks provide a robust and faithful assessment of model intelligence, revealing that there is still substantial room to improve in interactive scenarios. Project page: https://github.com/interactivebench/interactivebench</p> - https://arxiv.org/abs/2603.04737 - Thu, 05 Mar 2026 02:18:26 +0000 + Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights + https://arxiv.org/abs/2603.12228 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12228.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yulu Gan, Phillip Isola</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Pretraining produces a learned parameter vector that is typically treated as a starting point for further iterative adaptation. In this work, we instead view the outcome of pretraining as a distribution over parameter vectors, whose support already contains task-specific experts. We show that in small models such expert solutions occupy a negligible fraction of the volume of this distribution, making their discovery reliant on structured optimization methods such as gradient descent. In contrast, in large, well-pretrained models the density of task-experts increases dramatically, so that diverse, task-improving specialists populate a substantial fraction of the neighborhood around the pretrained weights. Motivated by this perspective, we explore a simple, fully parallel post-training method that samples N parameter perturbations at random, selects the top K, and ensembles predictions via majority vote. Despite its simplicity, this approach is competitive with standard post-training methods such as PPO, GRPO, and ES for contemporary large-scale models.</p> + https://arxiv.org/abs/2603.12228 + Thu, 12 Mar 2026 17:49:30 +0000 - DARE: Aligning LLM Agents with the R Statistical Ecosystem via Distribution-Aware Retrieval - https://arxiv.org/abs/2603.04743 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04743.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Maojun Sun, Yue Wu, Yifei Xie, Ruijian Han, Binyan Jiang, Defeng Sun, Yancheng Yuan, Jian Huang</p><p><b>Upvotes:</b> 44</p><p><b>Summary:</b> Large Language Model (LLM) agents can automate data-science workflows, but many rigorous statistical methods implemented in R remain underused because LLMs struggle with statistical knowledge and tool retrieval. Existing retrieval-augmented approaches focus on function-level semantics and ignore data distribution, producing suboptimal matches. We propose DARE (Distribution-Aware Retrieval Embedding), a lightweight, plug-and-play retrieval model that incorporates data distribution information into function representations for R package retrieval. Our main contributions are: (i) RPKB, a curated R Package Knowledge Base derived from 8,191 high-quality CRAN packages; (ii) DARE, an embedding model that fuses distributional features with function metadata to improve retrieval relevance; and (iii) RCodingAgent, an R-oriented LLM agent for reliable R code generation and a suite of statistical analysis tasks for systematically evaluating LLM agents in realistic analytical scenarios. Empirically, DARE achieves an NDCG at 10 of 93.47%, outperforming state-of-the-art open-source embedding models by up to 17% on package retrieval while using substantially fewer parameters. Integrating DARE into RCodingAgent yields significant gains on downstream analysis tasks. This work helps narrow the gap between LLM automation and the mature R statistical ecosystem.</p> - https://arxiv.org/abs/2603.04743 - Thu, 05 Mar 2026 02:29:54 +0000 + One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers + https://arxiv.org/abs/2603.12245 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12245.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Moayed Haji-Ali, Willi Menapace, Ivan Skorokhodov, Dogyun Park, Anil Kag, Michael Vasilkovsky, Sergey Tulyakov, Vicente Ordonez, Aliaksandr Siarohin</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Diffusion transformers (DiTs) achieve high generative quality but lock FLOPs to image resolution, limiting principled latency-quality trade-offs, and allocate computation uniformly across input spatial tokens, wasting resource allocation to unimportant regions. We introduce Elastic Latent Interface Transformer (ELIT), a drop-in, DiT-compatible mechanism that decouples input image size from compute. Our approach inserts a latent interface, a learnable variable-length token sequence on which standard transformer blocks can operate. Lightweight Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions. By training with random dropping of tail latents, ELIT learns to produce importance-ordered representations with earlier latents capturing global structure while later ones contain information to refine details. At inference, the number of latents can be dynamically adjusted to match compute constraints. ELIT is deliberately minimal, adding two cross-attention layers while leaving the rectified flow objective and the DiT stack unchanged. Across datasets and architectures (DiT, U-ViT, HDiT, MM-DiT), ELIT delivers consistent gains. On ImageNet-1K 512px, ELIT delivers an average gain of 35.3% and 39.6% in FID and FDD scores. Project page: https://snap-research.github.io/elit/</p> + https://arxiv.org/abs/2603.12245 + Thu, 12 Mar 2026 17:57:04 +0000 - Timer-S1: A Billion-Scale Time Series Foundation Model with Serial Scaling - https://arxiv.org/abs/2603.04791 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04791.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yong Liu, Xingjian Su, Shiyu Wang, Haoran Zhang, Haixuan Liu, Yuxuan Wang, Zhou Ye, Yang Xiang, Jianmin Wang, Mingsheng Long</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> We introduce Timer-S1, a strong Mixture-of-Experts (MoE) time series foundation model with 8.3B total parameters, 0.75B activated parameters for each token, and a context length of 11.5K. To overcome the scalability bottleneck in existing pre-trained time series foundation models, we perform Serial Scaling in three dimensions: model architecture, dataset, and training pipeline. Timer-S1 integrates sparse TimeMoE blocks and generic TimeSTP blocks for Serial-Token Prediction (STP), a generic training objective that adheres to the serial nature of forecasting. The proposed paradigm introduces serial computations to improve long-term predictions while avoiding costly rolling-style inference and pronounced error accumulation in the standard next-token prediction. Pursuing a high-quality and unbiased training dataset, we curate TimeBench, a corpus with one trillion time points, and apply meticulous data augmentation to mitigate predictive bias. We further pioneer a post-training stage, including continued pre-training and long-context extension, to enhance short-term and long-context performance. Evaluated on the large-scale GIFT-Eval leaderboard, Timer-S1 achieves state-of-the-art forecasting performance, attaining the best MASE and CRPS scores as a pre-trained model. Timer-S1 will be released to facilitate further research.</p> - https://arxiv.org/abs/2603.04791 - Thu, 05 Mar 2026 04:13:57 +0000 + Examining Reasoning LLMs-as-Judges in Non-Verifiable LLM Post-Training + https://arxiv.org/abs/2603.12246 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12246.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yixin Liu, Yue Yu, DiJia Su, Sid Wang, Xuewei Wang, Song Jiang, Bo Liu, Arman Cohan, Yuandong Tian, Zhengxing Chen</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Reasoning LLMs-as-Judges, which can benefit from inference-time scaling, provide a promising path for extending the success of reasoning models to non-verifiable domains where the output correctness/quality cannot be directly checked. However, while reasoning judges have shown better performance on static evaluation benchmarks, their effectiveness in actual policy training has not been systematically examined. Therefore, we conduct a rigorous study to investigate the actual impact of non-reasoning and reasoning judges in reinforcement-learning-based LLM alignment. Our controlled synthetic setting, where a "gold-standard" judge (gpt-oss-120b) provides preference annotations to train smaller judges, reveals key differences between non-reasoning and reasoning judges: non-reasoning judges lead to reward hacking easily, while reasoning judges can lead to policies that achieve strong performance when evaluated by the gold-standard judge. Interestingly, we find that the reasoning-judge-trained policies achieve such strong performance by learning to generate highly effective adversarial outputs that can also score well on popular benchmarks such as Arena-Hard by deceiving other LLM-judges. Combined with our further analysis, our study highlights both important findings and room for improvements for applying (reasoning) LLM-judges in non-verifiable LLM post-training.</p> + https://arxiv.org/abs/2603.12246 + Thu, 12 Mar 2026 17:57:06 +0000 - MASQuant: Modality-Aware Smoothing Quantization for Multimodal Large Language Models - https://arxiv.org/abs/2603.04800 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04800.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lulu Hu, Wenhu Xiao, Xin Chen, Xinhua Xu, Bowen Xu, Kun Li, Yongliang Tao</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Post-training quantization (PTQ) with computational invariance for Large Language Models~(LLMs) have demonstrated remarkable advances, however, their application to Multimodal Large Language Models~(MLLMs) presents substantial challenges. In this paper, we analyze SmoothQuant as a case study and identify two critical issues: Smoothing Misalignment and Cross-Modal Computational Invariance. To address these issues, we propose Modality-Aware Smoothing Quantization (MASQuant), a novel framework that introduces (1) Modality-Aware Smoothing (MAS), which learns separate, modality-specific smoothing factors to prevent Smoothing Misalignment, and (2) Cross-Modal Compensation (CMC), which addresses Cross-modal Computational Invariance by using SVD whitening to transform multi-modal activation differences into low-rank forms, enabling unified quantization across modalities. MASQuant demonstrates stable quantization performance across both dual-modal and tri-modal MLLMs. Experimental results show that MASQuant is competitive among the state-of-the-art PTQ algorithms. Source code: https://github.com/alibaba/EfficientAI.</p> - https://arxiv.org/abs/2603.04800 - Thu, 05 Mar 2026 04:41:32 +0000 + Trust Your Critic: Robust Reward Modeling and Reinforcement Learning for Faithful Image Editing and Generation + https://arxiv.org/abs/2603.12247 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12247.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiangyu Zhao, Peiyuan Zhang, Junming Lin, Tianhao Liang, Yuchen Duan, Shengyuan Ding, Changyao Tian, Yuhang Zang, Junchi Yan, Xue Yang</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> Reinforcement learning (RL) has emerged as a promising paradigm for enhancing image editing and text-to-image (T2I) generation. However, current reward models, which act as critics during RL, often suffer from hallucinations and assign noisy scores, inherently misguiding the optimization process. In this paper, we present FIRM (Faithful Image Reward Modeling), a comprehensive framework that develops robust reward models to provide accurate and reliable guidance for faithful image generation and editing. First, we design tailored data curation pipelines to construct high-quality scoring datasets. Specifically, we evaluate editing using both execution and consistency, while generation is primarily assessed via instruction following. Using these pipelines, we collect the FIRM-Edit-370K and FIRM-Gen-293K datasets, and train specialized reward models (FIRM-Edit-8B and FIRM-Gen-8B) that accurately reflect these criteria. Second, we introduce FIRM-Bench, a comprehensive benchmark specifically designed for editing and generation critics. Evaluations demonstrate that our models achieve superior alignment with human judgment compared to existing metrics. Furthermore, to seamlessly integrate these critics into the RL pipeline, we formulate a novel "Base-and-Bonus" reward strategy that balances competing objectives: Consistency-Modulated Execution (CME) for editing and Quality-Modulated Alignment (QMA) for generation. Empowered by this framework, our resulting models FIRM-Qwen-Edit and FIRM-SD3.5 achieve substantial performance breakthroughs. Comprehensive experiments demonstrate that FIRM mitigates hallucinations, establishing a new standard for fidelity and instruction adherence over existing general models. All of our datasets, models, and code have been publicly available at https://firm-reward.github.io.</p> + https://arxiv.org/abs/2603.12247 + Thu, 12 Mar 2026 17:57:21 +0000 - Locality-Attending Vision Transformer - https://arxiv.org/abs/2603.04892 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.04892.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sina Hajimiri, Farzad Beizaee, Fereshteh Shakeri, Christian Desrosiers, Ismail Ben Ayed, Jose Dolz</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Vision transformers have demonstrated remarkable success in classification by leveraging global self-attention to capture long-range dependencies. However, this same mechanism can obscure fine-grained spatial details crucial for tasks such as segmentation. In this work, we seek to enhance segmentation performance of vision transformers after standard image-level classification training. More specifically, we present a simple yet effective add-on that improves performance on segmentation tasks while retaining vision transformers' image-level recognition capabilities. In our approach, we modulate the self-attention with a learnable Gaussian kernel that biases the attention toward neighboring patches. We further refine the patch representations to learn better embeddings at patch positions. These modifications encourage tokens to focus on local surroundings and ensure meaningful representations at spatial positions, while still preserving the model's ability to incorporate global information. Experiments demonstrate the effectiveness of our modifications, evidenced by substantial segmentation gains on three benchmarks (e.g., over 6% and 4% on ADE20K for ViT Tiny and Base), without changing the training regime or sacrificing classification performance. The code is available at https://github.com/sinahmr/LocAtViT/.</p> - https://arxiv.org/abs/2603.04892 - Thu, 05 Mar 2026 07:31:49 +0000 + DVD: Deterministic Video Depth Estimation with Generative Priors + https://arxiv.org/abs/2603.12250 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12250.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hongfei Zhang, Harold Haodong Chen, Chenfei Liao, Jing He, Zixin Zhang, Haodong Li, Yihao Liang, Kanghao Chen, Bin Ren, Xu Zheng, Shuai Yang, Kun Zhou, Yinchuan Li, Nicu Sebe, Ying-Cong Chen</p><p><b>Upvotes:</b> 18</p><p><b>Summary:</b> Existing video depth estimation faces a fundamental trade-off: generative models suffer from stochastic geometric hallucinations and scale drift, while discriminative models demand massive labeled datasets to resolve semantic ambiguities. To break this impasse, we present DVD, the first framework to deterministically adapt pre-trained video diffusion models into single-pass depth regressors. Specifically, DVD features three core designs: (i) repurposing the diffusion timestep as a structural anchor to balance global stability with high-frequency details; (ii) latent manifold rectification (LMR) to mitigate regression-induced over-smoothing, enforcing differential constraints to restore sharp boundaries and coherent motion; and (iii) global affine coherence, an inherent property bounding inter-window divergence, which enables seamless long-video inference without requiring complex temporal alignment. Extensive experiments demonstrate that DVD achieves state-of-the-art zero-shot performance across benchmarks. Furthermore, DVD successfully unlocks the profound geometric priors implicit in video foundation models using 163x less task-specific data than leading baselines. Notably, we fully release our pipeline, providing the whole training suite for SOTA video depth estimation to benefit the open-source community.</p> + https://arxiv.org/abs/2603.12250 + Thu, 12 Mar 2026 17:58:06 +0000 - KARL: Knowledge Agents via Reinforcement Learning - https://arxiv.org/abs/2603.05218 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05218.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jonathan D. Chang, Andrew Drozdov, Shubham Toshniwal, Owen Oertell, Alexander Trott, Jacob Portes, Abhay Gupta, Pallavi Koppol, Ashutosh Baheti, Sean Kulinski, Ivan Zhou, Irene Dea, Krista Opsahl-Ong, Simon Favreau-Lessard, Sean Owen, Jose Javier Gonzalez Ortiz, Arnav Singhvi, Xabi Andrade, Cindy Wang, Kartik Sreenivasan, Sam Havens, Jialu Liu, Peyton DeNiro, Wen Sun, Michael Bendersky, Jonathan Frankle</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> We present a system for training enterprise search agents via reinforcement learning that achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks. Our work makes four core contributions. First, we introduce KARLBench, a multi-capability evaluation suite spanning six distinct search regimes, including constraint-driven entity search, cross-document report synthesis, tabular numerical reasoning, exhaustive entity retrieval, procedural reasoning over technical documentation, and fact aggregation over internal enterprise notes. Second, we show that models trained across heterogeneous search behaviors generalize substantially better than those optimized for any single benchmark. Third, we develop an agentic synthesis pipeline that employs long-horizon reasoning and tool use to generate diverse, grounded, and high-quality training data, with iterative bootstrapping from increasingly capable models. Fourth, we propose a new post-training paradigm based on iterative large-batch off-policy RL that is sample efficient, robust to train-inference engine discrepancies, and naturally extends to multi-task training with out-of-distribution generalization. Compared to Claude 4.6 and GPT 5.2, KARL is Pareto-optimal on KARLBench across cost-quality and latency-quality trade-offs, including tasks that were out-of-distribution during training. With sufficient test-time compute, it surpasses the strongest closed models. These results show that tailored synthetic data in combination with multi-task reinforcement learning enables cost-efficient and high-performing knowledge agents for grounded reasoning.</p> - https://arxiv.org/abs/2603.05218 - Thu, 05 Mar 2026 14:30:25 +0000 + EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models + https://arxiv.org/abs/2603.12252 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12252.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xuanlang Dai, Yujie Zhou, Long Xing, Jiazi Bu, Xilin Wei, Yuhong Liu, Beichen Zhang, Kai Chen, Yuhang Zang</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Recently, Multimodal Large Language Models (MLLMs) have been widely integrated into diffusion frameworks primarily as text encoders to tackle complex tasks such as spatial reasoning. However, this paradigm suffers from two critical limitations: (i) MLLMs text encoder exhibits insufficient reasoning depth. Single-step encoding fails to activate the Chain-of-Thought process, which is essential for MLLMs to provide accurate guidance for complex tasks. (ii) The guidance remains invariant during the decoding process. Invariant guidance during decoding prevents DiT from progressively decomposing complex instructions into actionable denoising steps, even with correct MLLM encodings. To this end, we propose Endogenous Chain-of-Thought (EndoCoT), a novel framework that first activates MLLMs' reasoning potential by iteratively refining latent thought states through an iterative thought guidance module, and then bridges these states to the DiT's denoising process. Second, a terminal thought grounding module is applied to ensure the reasoning trajectory remains grounded in textual supervision by aligning the final state with ground-truth answers. With these two components, the MLLM text encoder delivers meticulously reasoned guidance, enabling the DiT to execute it progressively and ultimately solve complex tasks in a step-by-step manner. Extensive evaluations across diverse benchmarks (e.g., Maze, TSP, VSP, and Sudoku) achieve an average accuracy of 92.1%, outperforming the strongest baseline by 8.3 percentage points.</p> + https://arxiv.org/abs/2603.12252 + Thu, 12 Mar 2026 17:58:48 +0000 - UltraDexGrasp: Learning Universal Dexterous Grasping for Bimanual Robots with Synthetic Data - https://arxiv.org/abs/2603.05312 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05312.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sizhe Yang, Yiman Xie, Zhixuan Liang, Yang Tian, Jia Zeng, Dahua Lin, Jiangmiao Pang</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Grasping is a fundamental capability for robots to interact with the physical world. Humans, equipped with two hands, autonomously select appropriate grasp strategies based on the shape, size, and weight of objects, enabling robust grasping and subsequent manipulation. In contrast, current robotic grasping remains limited, particularly in multi-strategy settings. Although substantial efforts have targeted parallel-gripper and single-hand grasping, dexterous grasping for bimanual robots remains underexplored, with data being a primary bottleneck. Achieving physically plausible and geometrically conforming grasps that can withstand external wrenches poses significant challenges. To address these issues, we introduce UltraDexGrasp, a framework for universal dexterous grasping with bimanual robots. The proposed data-generation pipeline integrates optimization-based grasp synthesis with planning-based demonstration generation, yielding high-quality and diverse trajectories across multiple grasp strategies. With this framework, we curate UltraDexGrasp-20M, a large-scale, multi-strategy grasp dataset comprising 20 million frames across 1,000 objects. Based on UltraDexGrasp-20M, we further develop a simple yet effective grasp policy that takes point clouds as input, aggregates scene features via unidirectional attention, and predicts control commands. Trained exclusively on synthetic data, the policy achieves robust zero-shot sim-to-real transfer and consistently succeeds on novel objects with varied shapes, sizes, and weights, attaining an average success rate of 81.2% in real-world universal dexterous grasping. To facilitate future research on grasping with bimanual robots, we open-source the data generation pipeline at https://github.com/InternRobotics/UltraDexGrasp.</p> - https://arxiv.org/abs/2603.05312 - Thu, 05 Mar 2026 15:54:21 +0000 + Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training + https://arxiv.org/abs/2603.12255 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12255.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Fangfu Liu, Diankun Wu, Jiawei Chi, Yimo Cai, Yi-Hsin Hung, Xumin Yu, Hao Li, Han Hu, Yongming Rao, Yueqi Duan</p><p><b>Upvotes:</b> 73</p><p><b>Summary:</b> Humans perceive and understand real-world spaces through a stream of visual observations. Therefore, the ability to streamingly maintain and update spatial evidence from potentially unbounded video streams is essential for spatial intelligence. The core challenge is not simply longer context windows but how spatial information is selected, organized, and retained over time. In this paper, we propose Spatial-TTT towards streaming visual-based spatial intelligence with test-time training (TTT), which adapts a subset of parameters (fast weights) to capture and organize spatial evidence over long-horizon scene videos. Specifically, we design a hybrid architecture and adopt large-chunk updates parallel with sliding-window attention for efficient spatial video processing. To further promote spatial awareness, we introduce a spatial-predictive mechanism applied to TTT layers with 3D spatiotemporal convolution, which encourages the model to capture geometric correspondence and temporal continuity across frames. Beyond architecture design, we construct a dataset with dense 3D spatial descriptions, which guides the model to update its fast weights to memorize and organize global 3D spatial signals in a structured manner. Extensive experiments demonstrate that Spatial-TTT improves long-horizon spatial understanding and achieves state-of-the-art performance on video spatial benchmarks. Project page: https://liuff19.github.io/Spatial-TTT.</p> + https://arxiv.org/abs/2603.12255 + Thu, 12 Mar 2026 17:58:58 +0000 - On-Policy Self-Distillation for Reasoning Compression - https://arxiv.org/abs/2603.05433 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05433.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hejian Sang, Yuanda Xu, Zhengze Zhou, Ran He, Zhipeng Wang, Jiachen Sun</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Reasoning models think out loud, but much of what they say is noise. We introduce OPSDC (On-Policy Self-Distillation for Reasoning Compression), a method that teaches models to reason more concisely by distilling their own concise behavior back into themselves. The entire approach reduces to one idea: condition the same model on a "be concise" instruction to obtain teacher logits, and minimize per-token reverse KL on the student's own rollouts. No ground-truth answers, no token budgets, no difficulty estimators. Just self-distillation. Yet this simplicity belies surprising sophistication: OPSDC automatically compresses easy problems aggressively while preserving the deliberation needed for hard ones. On Qwen3-8B and Qwen3-14B, we achieve 57-59% token reduction on MATH-500 while improving accuracy by 9-16 points absolute. On AIME 2024, the 14B model gains 10 points with 41% compression. The secret? Much of what reasoning models produce is not just redundant-it is actively harmful, compounding errors with every unnecessary token.</p> - https://arxiv.org/abs/2603.05433 - Thu, 05 Mar 2026 17:54:40 +0000 + DreamVideo-Omni: Omni-Motion Controlled Multi-Subject Video Customization with Latent Identity Reinforcement Learning + https://arxiv.org/abs/2603.12257 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12257.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yujie Wei, Xinyu Liu, Shiwei Zhang, Hangjie Yuan, Jinbo Xing, Zhekai Chen, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Ruihang Chu, Yingya Zhang, Yike Guo, Xihui Liu, Hongming Shan</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> While large-scale diffusion models have revolutionized video synthesis, achieving precise control over both multi-subject identity and multi-granularity motion remains a significant challenge. Recent attempts to bridge this gap often suffer from limited motion granularity, control ambiguity, and identity degradation, leading to suboptimal performance on identity preservation and motion control. In this work, we present DreamVideo-Omni, a unified framework enabling harmonious multi-subject customization with omni-motion control via a progressive two-stage training paradigm. In the first stage, we integrate comprehensive control signals for joint training, encompassing subject appearances, global motion, local dynamics, and camera movements. To ensure robust and precise controllability, we introduce a condition-aware 3D rotary positional embedding to coordinate heterogeneous inputs and a hierarchical motion injection strategy to enhance global motion guidance. Furthermore, to resolve multi-subject ambiguity, we introduce group and role embeddings to explicitly anchor motion signals to specific identities, effectively disentangling complex scenes into independent controllable instances. In the second stage, to mitigate identity degradation, we design a latent identity reward feedback learning paradigm by training a latent identity reward model upon a pretrained video diffusion backbone. This provides motion-aware identity rewards in the latent space, prioritizing identity preservation aligned with human preferences. Supported by our curated large-scale dataset and the comprehensive DreamOmni Bench for multi-subject and omni-motion control evaluation, DreamVideo-Omni demonstrates superior performance in generating high-quality videos with precise controllability.</p> + https://arxiv.org/abs/2603.12257 + Thu, 12 Mar 2026 17:59:12 +0000 - RealWonder: Real-Time Physical Action-Conditioned Video Generation - https://arxiv.org/abs/2603.05449 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05449.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wei Liu, Ziyu Chen, Zizhang Li, Yue Wang, Hong-Xing Yu, Jiajun Wu</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Current video generation models cannot simulate physical consequences of 3D actions like forces and robotic manipulations, as they lack structural understanding of how actions affect 3D scenes. We present RealWonder, the first real-time system for action-conditioned video generation from a single image. Our key insight is using physics simulation as an intermediate bridge: instead of directly encoding continuous actions, we translate them through physics simulation into visual representations (optical flow and RGB) that video models can process. RealWonder integrates three components: 3D reconstruction from single images, physics simulation, and a distilled video generator requiring only 4 diffusion steps. Our system achieves 13.2 FPS at 480x832 resolution, enabling interactive exploration of forces, robot actions, and camera controls on rigid objects, deformable bodies, fluids, and granular materials. We envision RealWonder opens new opportunities to apply video models in immersive experiences, AR/VR, and robot learning. Our code and model weights are publicly available in our project website: https://liuwei283.github.io/RealWonder/</p> - https://arxiv.org/abs/2603.05449 - Thu, 05 Mar 2026 18:22:54 +0000 + GRADE: Benchmarking Discipline-Informed Reasoning in Image Editing + https://arxiv.org/abs/2603.12264 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12264.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Mingxin Liu, Ziqian Fan, Zhaokai Wang, Leyao Gu, Zirun Zhu, Yiguo He, Yuchen Yang, Changyao Tian, Xiangyu Zhao, Ning Liao, Shaofeng Zhang, Qibing Ren, Zhihang Zhong, Xuanhe Zhou, Junchi Yan, Xue Yang</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Unified multimodal models target joint understanding, reasoning, and generation, but current image editing benchmarks are largely confined to natural images and shallow commonsense reasoning, offering limited assessment of this capability under structured, domain-specific constraints. In this work, we introduce GRADE, the first benchmark to assess discipline-informed knowledge and reasoning in image editing. GRADE comprises 520 carefully curated samples across 10 academic domains, spanning from natural science to social science. To support rigorous evaluation, we propose a multi-dimensional evaluation protocol that jointly assesses Discipline Reasoning, Visual Consistency, and Logical Readability. Extensive experiments on 20 state-of-the-art open-source and closed-source models reveal substantial limitations in current models under implicit, knowledge-intensive editing settings, leading to large performance gaps. Beyond quantitative scores, we conduct rigorous analyses and ablations to expose model shortcomings and identify the constraints within disciplinary editing. Together, GRADE pinpoints key directions for the future development of unified multimodal models, advancing the research on discipline-informed image editing and reasoning. Our benchmark and evaluation code are publicly released.</p> + https://arxiv.org/abs/2603.12264 + Thu, 12 Mar 2026 17:59:52 +0000 - Towards Multimodal Lifelong Understanding: A Dataset and Agentic Baseline - https://arxiv.org/abs/2603.05484 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05484.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Guo Chen, Lidong Lu, Yicheng Liu, Liangrui Dong, Lidong Zou, Jixin Lv, Zhenquan Li, Xinyi Mao, Baoqi Pei, Shihao Wang, Zhiqi Li, Karan Sapra, Fuxiao Liu, Yin-Dong Zheng, Yifei Huang, Limin Wang, Zhiding Yu, Andrew Tao, Guilin Liu, Tong Lu</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> While datasets for video understanding have scaled to hour-long durations, they typically consist of densely concatenated clips that differ from natural, unscripted daily life. To bridge this gap, we introduce MM-Lifelong, a dataset designed for Multimodal Lifelong Understanding. Comprising 181.1 hours of footage, it is structured across Day, Week, and Month scales to capture varying temporal densities. Extensive evaluations reveal two critical failure modes in current paradigms: end-to-end MLLMs suffer from a Working Memory Bottleneck due to context saturation, while representative agentic baselines experience Global Localization Collapse when navigating sparse, month-long timelines. To address this, we propose the Recursive Multimodal Agent (ReMA), which employs dynamic memory management to iteratively update a recursive belief state, significantly outperforming existing methods. Finally, we establish dataset splits designed to isolate temporal and domain biases, providing a rigorous foundation for future research in supervised learning and out-of-distribution generalization.</p> - https://arxiv.org/abs/2603.05484 - Thu, 05 Mar 2026 18:52:12 +0000 + OmniStream: Mastering Perception, Reconstruction and Action in Continuous Streams + https://arxiv.org/abs/2603.12265 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12265.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yibin Yan, Jilan Xu, Shangzhe Di, Haoning Wu, Weidi Xie</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Modern visual agents require representations that are general, causal, and physically structured to operate in real-time streaming environments. However, current vision foundation models remain fragmented, specializing narrowly in image semantic perception, offline temporal modeling, or spatial geometry. This paper introduces OmniStream, a unified streaming visual backbone that effectively perceives, reconstructs, and acts from diverse visual inputs. By incorporating causal spatiotemporal attention and 3D rotary positional embeddings (3D-RoPE), our model supports efficient, frame-by-frame online processing of video streams via a persistent KV-cache. We pre-train OmniStream using a synergistic multi-task framework coupling static and temporal representation learning, streaming geometric reconstruction, and vision-language alignment on 29 datasets. Extensive evaluations show that, even with a strictly frozen backbone, OmniStream achieves consistently competitive performance with specialized experts across image and video probing, streaming geometric reconstruction, complex video and spatial reasoning, as well as robotic manipulation (unseen at training). Rather than pursuing benchmark-specific dominance, our work demonstrates the viability of training a single, versatile vision backbone that generalizes across semantic, spatial, and temporal reasoning, i.e., a more meaningful step toward general-purpose visual understanding for interactive and embodied agents.</p> + https://arxiv.org/abs/2603.12265 + Thu, 12 Mar 2026 17:59:55 +0000 - RoboPocket: Improve Robot Policies Instantly with Your Phone - https://arxiv.org/abs/2603.05504 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.05504.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junjie Fang, Wendi Chen, Han Xue, Fangyuan Zhou, Tian Le, Yi Wang, Yuting Zhang, Jun Lv, Chuan Wen, Cewu Lu</p><p><b>Upvotes:</b> 30</p><p><b>Summary:</b> Scaling imitation learning is fundamentally constrained by the efficiency of data collection. While handheld interfaces have emerged as a scalable solution for in-the-wild data acquisition, they predominantly operate in an open-loop manner: operators blindly collect demonstrations without knowing the underlying policy's weaknesses, leading to inefficient coverage of critical state distributions. Conversely, interactive methods like DAgger effectively address covariate shift but rely on physical robot execution, which is costly and difficult to scale. To reconcile this trade-off, we introduce RoboPocket, a portable system that enables Robot-Free Instant Policy Iteration using single consumer smartphones. Its core innovation is a Remote Inference framework that visualizes the policy's predicted trajectory via Augmented Reality (AR) Visual Foresight. This immersive feedback allows collectors to proactively identify potential failures and focus data collection on the policy's weak regions without requiring a physical robot. Furthermore, we implement an asynchronous Online Finetuning pipeline that continuously updates the policy with incoming data, effectively closing the learning loop in minutes. Extensive experiments demonstrate that RoboPocket adheres to data scaling laws and doubles the data efficiency compared to offline scaling strategies, overcoming their long-standing efficiency bottleneck. Moreover, our instant iteration loop also boosts sample efficiency by up to 2times in distributed environments a small number of interactive corrections per person. Project page and videos: https://robo-pocket.github.io.</p> - https://arxiv.org/abs/2603.05504 - Thu, 05 Mar 2026 18:59:38 +0000 + EVATok: Adaptive Length Video Tokenization for Efficient Visual Autoregressive Generation + https://arxiv.org/abs/2603.12267 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2603.12267.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tianwei Xiong, Jun Hao Liew, Zilong Huang, Zhijie Lin, Jiashi Feng, Xihui Liu</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Autoregressive (AR) video generative models rely on video tokenizers that compress pixels into discrete token sequences. The length of these token sequences is crucial for balancing reconstruction quality against downstream generation computational cost. Traditional video tokenizers apply a uniform token assignment across temporal blocks of different videos, often wasting tokens on simple, static, or repetitive segments while underserving dynamic or complex ones. To address this inefficiency, we introduce EVATok, a framework to produce Efficient Video Adaptive Tokenizers. Our framework estimates optimal token assignments for each video to achieve the best quality-cost trade-off, develops lightweight routers for fast prediction of these optimal assignments, and trains adaptive tokenizers that encode videos based on the assignments predicted by routers. We demonstrate that EVATok delivers substantial improvements in efficiency and overall quality for video reconstruction and downstream AR generation. Enhanced by our advanced training recipe that integrates video semantic encoders, EVATok achieves superior reconstruction and state-of-the-art class-to-video generation on UCF-101, with at least 24.4% savings in average token usage compared to the prior state-of-the-art LARP and our fixed-length baseline.</p> + https://arxiv.org/abs/2603.12267 + Thu, 12 Mar 2026 17:59:59 +0000