From 6bb39932c1953fd00f99f900f0154440ac4f9122 Mon Sep 17 00:00:00 2001 From: "github-actions[bot]" Date: Sat, 14 Feb 2026 00:05:28 +0000 Subject: [PATCH] bot: update RSS feed --- feed.xml | 437 +++++++++++++++++++++++++------------------------------ 1 file changed, 201 insertions(+), 236 deletions(-) diff --git a/feed.xml b/feed.xml index c1b0f49..28ad42f 100644 --- a/feed.xml +++ b/feed.xml @@ -7,321 +7,286 @@ http://www.rssboard.org/rss-specification python-feedgen en - Fri, 13 Feb 2026 00:05:51 +0000 + Sat, 14 Feb 2026 00:05:28 +0000 - ECHO-2: A Large-Scale Distributed Rollout Framework for Cost-Efficient Reinforcement Learning - https://arxiv.org/abs/2602.02192 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.02192.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jie Xiao, Meng Chen, Qingnan Ren, Jingwei Song, Jiaqi Huang, Yangshen Deng, Chris Tong, Wanyi Chen, Suli Wang, Ziqian Bi, Shuo Lu, Yiqun Duan, Xu Wang, Rymon Yu, Ween Yang, Lynn Ai, Eric Yang, Bill Shi, Song Jingwei</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Reinforcement learning (RL) is a critical stage in post-training large language models (LLMs), involving repeated interaction between rollout generation, reward evaluation, and centralized learning. Distributing rollout execution offers opportunities to leverage more cost-efficient inference resources, but introduces challenges in wide-area coordination and policy dissemination. We present ECHO-2, a distributed RL framework for post-training with remote inference workers and non-negligible dissemination latency. ECHO-2 combines centralized learning with distributed rollouts and treats bounded policy staleness as a user-controlled parameter, enabling rollout generation, dissemination, and training to overlap. We introduce an overlap-based capacity model that relates training time, dissemination latency, and rollout throughput, yielding a practical provisioning rule for sustaining learner utilization. To mitigate dissemination bottlenecks and lower cost, ECHO-2 employs peer-assisted pipelined broadcast and cost-aware activation of heterogeneous workers. Experiments on GRPO post-training of 4B and 8B models under real wide-area bandwidth regimes show that ECHO-2 significantly improves cost efficiency while preserving RL reward comparable to strong baselines.</p> - https://arxiv.org/abs/2602.02192 - Mon, 02 Feb 2026 14:57:53 +0000 + Unveiling Implicit Advantage Symmetry: Why GRPO Struggles with Exploration and Difficulty Adaptation + https://arxiv.org/abs/2602.05548 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.05548.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhiqi Yu, Zhangquan Chen, Mengting Liu, Heye Zhang, Liangqiong Qu</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Reinforcement Learning with Verifiable Rewards (RLVR), particularly GRPO, has become the standard for eliciting LLM reasoning. However, its efficiency in exploration and difficulty adaptation remains an open challenge. In this work, we argue that these bottlenecks stem from an implicit advantage symmetry inherent in Group Relative Advantage Estimation (GRAE). This symmetry induces two critical limitations: (i) at the group level, strict symmetry in weights between correct and incorrect trajectories leaves unsampled action logits unchanged, thereby hindering exploration of novel correct solution. (ii) at the sample level, the algorithm implicitly prioritizes medium-difficulty samples, remaining agnostic to the non-stationary demands of difficulty focus. Through controlled experiments, we reveal that this symmetric property is sub-optimal, yielding two pivotal insights: (i) asymmetrically suppressing the advantages of correct trajectories encourages essential exploration. (ii) learning efficiency is maximized by a curriculum-like transition-prioritizing simpler samples initially before gradually shifting to complex ones. Motivated by these findings, we propose Asymmetric GRAE (A-GRAE), which dynamically modulates exploration incentives and sample-difficulty focus. Experiments across seven benchmarks demonstrate that A-GRAE consistently improves GRPO and its variants across both LLMs and MLLMs.</p> + https://arxiv.org/abs/2602.05548 + Thu, 05 Feb 2026 11:07:14 +0000 - TIC-VLA: A Think-in-Control Vision-Language-Action Model for Robot Navigation in Dynamic Environments - https://arxiv.org/abs/2602.02459 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.02459.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhiyu Huang, Yun Zhang, Johnson Liu, Rui Song, Chen Tang, Jiaqi Ma</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Robots in dynamic, human-centric environments must follow language instructions while maintaining real-time reactive control. Vision-language-action (VLA) models offer a promising framework, but they assume temporally aligned reasoning and control, despite semantic inference being inherently delayed relative to real-time action. We introduce Think-in-Control (TIC)-VLA, a latency-aware framework that explicitly models delayed semantic reasoning during action generation. TIC-VLA defines a delayed semantic-control interface that conditions action generation on delayed vision-language semantic states and explicit latency metadata, in addition to current observations, enabling policies to compensate for asynchronous reasoning. We further propose a latency-consistent training pipeline that injects reasoning inference delays during imitation learning and online reinforcement learning, aligning training with asynchronous deployment. To support realistic evaluation, we present DynaNav, a physics-accurate, photo-realistic simulation suite for language-guided navigation in dynamic environments. Extensive experiments in simulation and on a real robot show that TIC-VLA consistently outperforms prior VLA models while maintaining robust real-time control under multi-second reasoning latency. Project website: https://ucla-mobility.github.io/TIC-VLA/</p> - https://arxiv.org/abs/2602.02459 - Mon, 02 Feb 2026 18:47:49 +0000 + Sparse Video Generation Propels Real-World Beyond-the-View Vision-Language Navigation + https://arxiv.org/abs/2602.05827 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.05827.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hai Zhang, Siqi Liang, Li Chen, Yuxian Li, Yukuan Xu, Yichao Zhong, Fu Zhang, Hongyang Li</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Why must vision-language navigation be bound to detailed and verbose language instructions? While such details ease decision-making, they fundamentally contradict the goal for navigation in the real-world. Ideally, agents should possess the autonomy to navigate in unknown environments guided solely by simple and high-level intents. Realizing this ambition introduces a formidable challenge: Beyond-the-View Navigation (BVN), where agents must locate distant, unseen targets without dense and step-by-step guidance. Existing large language model (LLM)-based methods, though adept at following dense instructions, often suffer from short-sighted behaviors due to their reliance on short-horimzon supervision. Simply extending the supervision horizon, however, destabilizes LLM training. In this work, we identify that video generation models inherently benefit from long-horizon supervision to align with language instructions, rendering them uniquely suitable for BVN tasks. Capitalizing on this insight, we propose introducing the video generation model into this field for the first time. Yet, the prohibitive latency for generating videos spanning tens of seconds makes real-world deployment impractical. To bridge this gap, we propose SparseVideoNav, achieving sub-second trajectory inference guided by a generated sparse future spanning a 20-second horizon. This yields a remarkable 27x speed-up compared to the unoptimized counterpart. Extensive real-world zero-shot experiments demonstrate that SparseVideoNav achieves 2.5x the success rate of state-of-the-art LLM baselines on BVN tasks and marks the first realization of such capability in challenging night scenes.</p> + https://arxiv.org/abs/2602.05827 + Thu, 05 Feb 2026 16:16:13 +0000 - Reasoning Cache: Continual Improvement Over Long Horizons via Short-Horizon RL - https://arxiv.org/abs/2602.03773 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.03773.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ian Wu, Yuxiao Qu, Amrith Setlur, Aviral Kumar</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Large Language Models (LLMs) that can continually improve beyond their training budgets are able to solve increasingly difficult problems by adapting at test time, a property we refer to as extrapolation. However, standard reinforcement learning (RL) operates over fixed problem distributions and training budgets, which limits extrapolation amidst distribution shift at test time. To address this, we introduce RC, an iterative decoding algorithm that replaces standard autoregressive decoding during both training and inference. RC exploits an asymmetry between the response generation and summarization capabilities of LLMs to construct reasoning chains that consistently improve across iterations. Models trained to use RC can extrapolate and continually improve over reasoning horizons more than an order of magnitude longer than those seen during training. Empirically, training a 4B model with RC using a 16k-token training budget improves performance on HMMT 2025 from 40% to nearly 70% with 0.5m tokens at test time, outperforming both comparably sized models and many larger reasoning LLMs. Finally, we also show that models trained with RC can more effectively leverage existing scaffolds to further scale test-time performance, due to the improved summary-conditioned generation abilities learned through training.</p> - https://arxiv.org/abs/2602.03773 - Tue, 03 Feb 2026 17:34:04 +0000 + Dreaming in Code for Curriculum Learning in Open-Ended Worlds + https://arxiv.org/abs/2602.08194 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08194.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Konstantinos Mitsides, Maxence Faldor, Antoine Cully</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Open-ended learning frames intelligence as emerging from continual interaction with an ever-expanding space of environments. While recent advances have utilized foundation models to programmatically generate diverse environments, these approaches often focus on discovering isolated behaviors rather than orchestrating sustained progression. In complex open-ended worlds, the large combinatorial space of possible challenges makes it difficult for agents to discover sequences of experiences that remain consistently learnable. To address this, we propose Dreaming in Code (DiCode), a framework in which foundation models synthesize executable environment code to scaffold learning toward increasing competence. In DiCode, "dreaming" takes the form of materializing code-level variations of the world. We instantiate DiCode in Craftax, a challenging open-ended benchmark characterized by rich mechanics and long-horizon progression. Empirically, DiCode enables agents to acquire long-horizon skills, achieving a 16% improvement in mean return over the strongest baseline and non-zero success on late-game combat tasks where prior methods fail. Our results suggest that code-level environment design provides a practical mechanism for curriculum control, enabling the construction of intermediate environments that bridge competence gaps in open-ended worlds. Project page and source code are available at https://konstantinosmitsides.github.io/dreaming-in-code and https://github.com/konstantinosmitsides/dreaming-in-code.</p> + https://arxiv.org/abs/2602.08194 + Mon, 09 Feb 2026 01:24:40 +0000 - ASA: Training-Free Representation Engineering for Tool-Calling Agents - https://arxiv.org/abs/2602.04935 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.04935.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Youjin Wang, Run Zhou, Rong Fu, Shuaishuai Cao, Hongwei Zeng, Jiaxuan Lu, Sicheng Fan, Jiaqiao Zhao, Liangming Pan</p><p><b>Upvotes:</b> 39</p><p><b>Summary:</b> Adapting LLM agents to domain-specific tool calling remains notably brittle under evolving interfaces. Prompt and schema engineering is easy to deploy but often fragile under distribution shift and strict parsers, while continual parameter-efficient fine-tuning improves reliability at the cost of training, maintenance, and potential forgetting. We identify a critical Lazy Agent failure mode where tool necessity is nearly perfectly decodable from mid-layer activations, yet the model remains conservative in entering tool mode, revealing a representation-behavior gap. We propose Activation Steering Adapter (ASA), a training-free, inference-time controller that performs a single-shot mid-layer intervention and targets tool domains via a router-conditioned mixture of steering vectors with a probe-guided signed gate to amplify true intent while suppressing spurious triggers. On MTU-Bench with Qwen2.5-1.5B, ASA improves strict tool-use F1 from 0.18 to 0.50 while reducing the false positive rate from 0.15 to 0.05, using only about 20KB of portable assets and no weight updates.</p> - https://arxiv.org/abs/2602.04935 - Wed, 04 Feb 2026 14:20:02 +0000 + PISCO: Precise Video Instance Insertion with Sparse Control + https://arxiv.org/abs/2602.08277 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08277.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiangbo Gao, Renjie Li, Xinghao Chen, Yuheng Wu, Suofei Feng, Qing Yin, Zhengzhong Tu</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> The landscape of AI video generation is undergoing a pivotal shift: moving beyond general generation - which relies on exhaustive prompt-engineering and "cherry-picking" - towards fine-grained, controllable generation and high-fidelity post-processing. In professional AI-assisted filmmaking, it is crucial to perform precise, targeted modifications. A cornerstone of this transition is video instance insertion, which requires inserting a specific instance into existing footage while maintaining scene integrity. Unlike traditional video editing, this task demands several requirements: precise spatial-temporal placement, physically consistent scene interaction, and the faithful preservation of original dynamics - all achieved under minimal user effort. In this paper, we propose PISCO, a video diffusion model for precise video instance insertion with arbitrary sparse keyframe control. PISCO allows users to specify a single keyframe, start-and-end keyframes, or sparse keyframes at arbitrary timestamps, and automatically propagates object appearance, motion, and interaction. To address the severe distribution shift induced by sparse conditioning in pretrained video diffusion models, we introduce Variable-Information Guidance for robust conditioning and Distribution-Preserving Temporal Masking to stabilize temporal generation, together with geometry-aware conditioning for realistic scene adaptation. We further construct PISCO-Bench, a benchmark with verified instance annotations and paired clean background videos, and evaluate performance using both reference-based and reference-free perceptual metrics. Experiments demonstrate that PISCO consistently outperforms strong inpainting and video editing baselines under sparse control, and exhibits clear, monotonic performance improvements as additional control signals are provided. Project page: xiangbogaobarry.github.io/PISCO.</p> + https://arxiv.org/abs/2602.08277 + Mon, 09 Feb 2026 05:15:39 +0000 - AgenticPay: A Multi-Agent LLM Negotiation System for Buyer-Seller Transactions - https://arxiv.org/abs/2602.06008 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.06008.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xianyang Liu, Shangding Gu, Dawn Song</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Large language model (LLM)-based agents are increasingly expected to negotiate, coordinate, and transact autonomously, yet existing benchmarks lack principled settings for evaluating language-mediated economic interaction among multiple agents. We introduce AgenticPay, a benchmark and simulation framework for multi-agent buyer-seller negotiation driven by natural language. AgenticPay models markets in which buyers and sellers possess private constraints and product-dependent valuations, and must reach agreements through multi-round linguistic negotiation rather than numeric bidding alone. The framework supports a diverse suite of over 110 tasks ranging from bilateral bargaining to many-to-many markets, with structured action extraction and metrics for feasibility, efficiency, and welfare. Benchmarking state-of-the-art proprietary and open-weight LLMs reveals substantial gaps in negotiation performance and highlights challenges in long-horizon strategic reasoning, establishing AgenticPay as a foundation for studying agentic commerce and language-based market interaction. Code and dataset are available at the link: https://github.com/SafeRL-Lab/AgenticPay.</p> - https://arxiv.org/abs/2602.06008 - Thu, 05 Feb 2026 18:50:36 +0000 + NarraScore: Bridging Visual Narrative and Musical Dynamics via Hierarchical Affective Control + https://arxiv.org/abs/2602.09070 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09070.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yufan Wen, Zhaocheng Liu, YeGuo Hua, Ziyi Guo, Lihua Zhang, Chun Yuan, Jian Wu</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Synthesizing coherent soundtracks for long-form videos remains a formidable challenge, currently stalled by three critical impediments: computational scalability, temporal coherence, and, most critically, a pervasive semantic blindness to evolving narrative logic. To bridge these gaps, we propose NarraScore, a hierarchical framework predicated on the core insight that emotion serves as a high-density compression of narrative logic. Uniquely, we repurpose frozen Vision-Language Models (VLMs) as continuous affective sensors, distilling high-dimensional visual streams into dense, narrative-aware Valence-Arousal trajectories. Mechanistically, NarraScore employs a Dual-Branch Injection strategy to reconcile global structure with local dynamism: a Global Semantic Anchor ensures stylistic stability, while a surgical Token-Level Affective Adapter modulates local tension via direct element-wise residual injection. This minimalist design bypasses the bottlenecks of dense attention and architectural cloning, effectively mitigating the overfitting risks associated with data scarcity. Experiments demonstrate that NarraScore achieves state-of-the-art consistency and narrative alignment with negligible computational overhead, establishing a fully autonomous paradigm for long-video soundtrack generation.</p> + https://arxiv.org/abs/2602.09070 + Mon, 09 Feb 2026 09:39:42 +0000 - Ex-Omni: Enabling 3D Facial Animation Generation for Omni-modal Large Language Models - https://arxiv.org/abs/2602.07106 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.07106.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haoyu Zhang, Zhipeng Li, Yiwen Guo, Tianshu Yu</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Omni-modal large language models (OLLMs) aim to unify multimodal understanding and generation, yet incorporating speech with 3D facial animation remains largely unexplored despite its importance for natural interaction. A key challenge arises from the representation mismatch between discrete, token-level semantic reasoning in LLMs and the dense, fine-grained temporal dynamics required for 3D facial motion, which makes direct modeling difficult to optimize under limited data. We propose Expressive Omni (Ex-Omni), an open-source omni-modal framework that augments OLLMs with speech-accompanied 3D facial animation. Ex-Omni reduces learning difficulty by decoupling semantic reasoning from temporal generation, leveraging speech units as temporal scaffolding and a unified token-as-query gated fusion (TQGF) mechanism for controlled semantic injection. We further introduce InstructEx, a dataset aims to facilitate augment OLLMs with speech-accompanied 3D facial animation. Extensive experiments demonstrate that Ex-Omni performs competitively against existing open-source OLLMs while enabling stable aligned speech and facial animation generation.</p> - https://arxiv.org/abs/2602.07106 - Fri, 06 Feb 2026 18:03:30 +0000 + χ_{0}: Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies + https://arxiv.org/abs/2602.09021 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09021.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Checheng Yu, Chonghao Sima, Gangcheng Jiang, Hai Zhang, Haoguang Mai, Hongyang Li, Huijie Wang, Jin Chen, Kaiyang Wu, Li Chen, Lirui Zhao, Modi Shi, Ping Luo, Qingwen Bu, Shijia Peng, Tianyu Li, Yibo Yuan</p><p><b>Upvotes:</b> 16</p><p><b>Summary:</b> High-reliability long-horizon robotic manipulation has traditionally relied on large-scale data and compute to understand complex real-world dynamics. However, we identify that the primary bottleneck to real-world robustness is not resource scale alone, but the distributional shift among the human demonstration distribution, the inductive bias learned by the policy, and the test-time execution distribution -- a systematic inconsistency that causes compounding errors in multi-stage tasks. To mitigate these inconsistencies, we propose χ_{0}, a resource-efficient framework with effective modules designated to achieve production-level robustness in robotic manipulation. Our approach builds off three technical pillars: (i) Model Arithmetic, a weight-space merging strategy that efficiently soaks up diverse distributions of different demonstrations, varying from object appearance to state variations; (ii) Stage Advantage, a stage-aware advantage estimator that provides stable, dense progress signals, overcoming the numerical instability of prior non-stage approaches; and (iii) Train-Deploy Alignment, which bridges the distribution gap via spatio-temporal augmentation, heuristic DAgger corrections, and temporal chunk-wise smoothing. χ_{0} enables two sets of dual-arm robots to collaboratively orchestrate long-horizon garment manipulation, spanning tasks from flattening, folding, to hanging different clothes. Our method exhibits high-reliability autonomy; we are able to run the system from arbitrary initial state for consecutive 24 hours non-stop. Experiments validate that χ_{0} surpasses the state-of-the-art π_{0.5} in success rate by nearly 250%, with only 20-hour data and 8 A100 GPUs. Code, data and models will be released to facilitate the community.</p> + https://arxiv.org/abs/2602.09021 + Mon, 09 Feb 2026 18:59:45 +0000 - Rethinking the Value of Agent-Generated Tests for LLM-Based Software Engineering Agents - https://arxiv.org/abs/2602.07900 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.07900.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhi Chen, Zhensu Sun, Yuling Shi, Chao Peng, Xiaodong Gu, David Lo, Lingxiao Jiang</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Large Language Model (LLM) code agents increasingly resolve repository-level issues by iteratively editing code, invoking tools, and validating candidate patches. In these workflows, agents often write tests on the fly, a paradigm adopted by many high-ranking agents on the SWE-bench leaderboard. However, we observe that GPT-5.2, which writes almost no new tests, can even achieve performance comparable to top-ranking agents. This raises the critical question: whether such tests meaningfully improve issue resolution or merely mimic human testing practices while consuming a substantial interaction budget. To reveal the impact of agent-written tests, we present an empirical study that analyzes agent trajectories across six state-of-the-art LLMs on SWE-bench Verified. Our results show that while test writing is commonly adopted, but resolved and unresolved tasks within the same model exhibit similar test-writing frequencies Furthermore, these tests typically serve as observational feedback channels, where agents prefer value-revealing print statements significantly more than formal assertion-based checks. Based on these insights, we perform a controlled experiment by revising the prompts of four agents to either increase or reduce test writing. The results suggest that changes in the volume of agent-written tests do not significantly change final outcomes. Taken together, our study reveals that current test-writing practices may provide marginal utility in autonomous software engineering tasks.</p> - https://arxiv.org/abs/2602.07900 - Sun, 08 Feb 2026 10:26:31 +0000 + The Devil Behind Moltbook: Anthropic Safety is Always Vanishing in Self-Evolving AI Societies + https://arxiv.org/abs/2602.09877 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09877.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chenxu Wang, Chaozhuo Li, Songyang Liu, Zejian Chen, Jinyu Hou, Ji Qi, Rui Li, Litian Zhang, Qiwei Ye, Zheng Liu, Xu Chen, Xi Zhang, Philip S. Yu</p><p><b>Upvotes:</b> 169</p><p><b>Summary:</b> The emergence of multi-agent systems built from large language models (LLMs) offers a promising paradigm for scalable collective intelligence and self-evolution. Ideally, such systems would achieve continuous self-improvement in a fully closed loop while maintaining robust safety alignment--a combination we term the self-evolution trilemma. However, we demonstrate both theoretically and empirically that an agent society satisfying continuous self-evolution, complete isolation, and safety invariance is impossible. Drawing on an information-theoretic framework, we formalize safety as the divergence degree from anthropic value distributions. We theoretically demonstrate that isolated self-evolution induces statistical blind spots, leading to the irreversible degradation of the system's safety alignment. Empirical and qualitative results from an open-ended agent community (Moltbook) and two closed self-evolving systems reveal phenomena that align with our theoretical prediction of inevitable safety erosion. We further propose several solution directions to alleviate the identified safety concern. Our work establishes a fundamental limit on the self-evolving AI societies and shifts the discourse from symptom-driven safety patches to a principled understanding of intrinsic dynamical risks, highlighting the need for external oversight or novel safety-preserving mechanisms.</p> + https://arxiv.org/abs/2602.09877 + Tue, 10 Feb 2026 15:18:19 +0000 - Bielik Guard: Efficient Polish Language Safety Classifiers for LLM Content Moderation - https://arxiv.org/abs/2602.07954 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.07954.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Krzysztof Wróbel, Jan Maria Kowalski, Jerzy Surma, Igor Ciuciura, Maciej Szymański</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> As Large Language Models (LLMs) become increasingly deployed in Polish language applications, the need for efficient and accurate content safety classifiers has become paramount. We present Bielik Guard, a family of compact Polish language safety classifiers comprising two model variants: a 0.1B parameter model based on MMLW-RoBERTa-base and a 0.5B parameter model based on PKOBP/polish-roberta-8k. Fine-tuned on a community-annotated dataset of 6,885 Polish texts, these models classify content across five safety categories: Hate/Aggression, Vulgarities, Sexual Content, Crime, and Self-Harm. Our evaluation demonstrates that both models achieve strong performance on multiple benchmarks. The 0.5B variant offers the best overall discrimination capability with F1 scores of 0.791 (micro) and 0.785 (macro) on the test set, while the 0.1B variant demonstrates exceptional efficiency. Notably, Bielik Guard 0.1B v1.1 achieves superior precision (77.65%) and very low false positive rate (0.63%) on real user prompts, outperforming HerBERT-PL-Guard (31.55% precision, 4.70% FPR) despite identical model size. The models are publicly available and designed to provide appropriate responses rather than simple content blocking, particularly for sensitive categories like self-harm.</p> - https://arxiv.org/abs/2602.07954 - Sun, 08 Feb 2026 12:57:04 +0000 + Stemphonic: All-at-once Flexible Multi-stem Music Generation + https://arxiv.org/abs/2602.09891 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09891.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shih-Lun Wu, Ge Zhu, Juan-Pablo Caceres, Cheng-Zhi Anna Huang, Nicholas J. Bryan</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Music stem generation, the task of producing musically-synchronized and isolated instrument audio clips, offers the potential of greater user control and better alignment with musician workflows compared to conventional text-to-music models. Existing stem generation approaches, however, either rely on fixed architectures that output a predefined set of stems in parallel, or generate only one stem at a time, resulting in slow inference despite flexibility in stem combination. We propose Stemphonic, a diffusion-/flow-based framework that overcomes this trade-off and generates a variable set of synchronized stems in one inference pass. During training, we treat each stem as a batch element, group synchronized stems in a batch, and apply a shared noise latent to each group. At inference-time, we use a shared initial noise latent and stem-specific text inputs to generate synchronized multi-stem outputs in one pass. We further expand our approach to enable one-pass conditional multi-stem generation and stem-wise activity controls to empower users to iteratively generate and orchestrate the temporal layering of a mix. We benchmark our results on multiple open-source stem evaluation sets and show that Stemphonic produces higher-quality outputs while accelerating the full mix generation process by 25 to 50%. Demos at: https://stemphonic-demo.vercel.app.</p> + https://arxiv.org/abs/2602.09891 + Tue, 10 Feb 2026 15:30:12 +0000 - Free(): Learning to Forget in Malloc-Only Reasoning Models - https://arxiv.org/abs/2602.08030 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08030.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yilun Zheng, Dongyang Ma, Tian Liang, Jiahao Xu, Xinting Huang, Lihui Chen, Haitao Mi, Yan Wang</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Reasoning models enhance problem-solving by scaling test-time compute, yet they face a critical paradox: excessive thinking tokens often degrade performance rather than improve it. We attribute this to a fundamental architectural flaw: standard LLMs operate as "malloc-only" engines, continuously accumulating valid and redundant steps alike without a mechanism to prune obsolete information. To break this cycle, we propose Free()LM, a model that introduces an intrinsic self-forgetting capability via the Free-Module, a plug-and-play LoRA adapter. By iteratively switching between reasoning and cleaning modes, Free()LM dynamically identifies and prunes useless context chunks, maintaining a compact and noise-free state. Extensive experiments show that Free()LM provides consistent improvements across all model scales (8B to 685B). It achieves a 3.3% average improvement over top-tier reasoning baselines, even establishing a new SOTA on IMOanswerBench using DeepSeek V3.2-Speciale. Most notably, in long-horizon tasks where the standard Qwen3-235B-A22B model suffers a total collapse (0% accuracy), Free()LM restores performance to 50%. Our findings suggest that sustainable intelligence requires the freedom to forget as much as the power to think.</p> - https://arxiv.org/abs/2602.08030 - Sun, 08 Feb 2026 16:04:23 +0000 + EgoHumanoid: Unlocking In-the-Wild Loco-Manipulation with Robot-Free Egocentric Demonstration + https://arxiv.org/abs/2602.10106 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10106.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Modi Shi, Shijia Peng, Jin Chen, Haoran Jiang, Yinghui Li, Di Huang, Ping Luo, Hongyang Li, Li Chen</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Human demonstrations offer rich environmental diversity and scale naturally, making them an appealing alternative to robot teleoperation. While this paradigm has advanced robot-arm manipulation, its potential for the more challenging, data-hungry problem of humanoid loco-manipulation remains largely unexplored. We present EgoHumanoid, the first framework to co-train a vision-language-action policy using abundant egocentric human demonstrations together with a limited amount of robot data, enabling humanoids to perform loco-manipulation across diverse real-world environments. To bridge the embodiment gap between humans and robots, including discrepancies in physical morphology and viewpoint, we introduce a systematic alignment pipeline spanning from hardware design to data processing. A portable system for scalable human data collection is developed, and we establish practical collection protocols to improve transferability. At the core of our human-to-humanoid alignment pipeline lies two key components. The view alignment reduces visual domain discrepancies caused by camera height and perspective variation. The action alignment maps human motions into a unified, kinematically feasible action space for humanoid control. Extensive real-world experiments demonstrate that incorporating robot-free egocentric data significantly outperforms robot-only baselines by 51\%, particularly in unseen environments. Our analysis further reveals which behaviors transfer effectively and the potential for scaling human data.</p> + https://arxiv.org/abs/2602.10106 + Tue, 10 Feb 2026 18:59:03 +0000 - Graph-Enhanced Deep Reinforcement Learning for Multi-Objective Unrelated Parallel Machine Scheduling - https://arxiv.org/abs/2602.08052 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08052.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Bulent Soykan, Sean Mondesire, Ghaith Rabadi, Grace Bochenek</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> The Unrelated Parallel Machine Scheduling Problem (UPMSP) with release dates, setups, and eligibility constraints presents a significant multi-objective challenge. Traditional methods struggle to balance minimizing Total Weighted Tardiness (TWT) and Total Setup Time (TST). This paper proposes a Deep Reinforcement Learning framework using Proximal Policy Optimization (PPO) and a Graph Neural Network (GNN). The GNN effectively represents the complex state of jobs, machines, and setups, allowing the PPO agent to learn a direct scheduling policy. Guided by a multi-objective reward function, the agent simultaneously minimizes TWT and TST. Experimental results on benchmark instances demonstrate that our PPO-GNN agent significantly outperforms a standard dispatching rule and a metaheuristic, achieving a superior trade-off between both objectives. This provides a robust and scalable solution for complex manufacturing scheduling.</p> - https://arxiv.org/abs/2602.08052 - Sun, 08 Feb 2026 16:54:47 +0000 + MetaphorStar: Image Metaphor Understanding and Reasoning with End-to-End Visual Reinforcement Learning + https://arxiv.org/abs/2602.10575 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10575.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chenhao Zhang, Yazhe Niu, Hongsheng Li</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Metaphorical comprehension in images remains a critical challenge for Nowadays AI systems. While Multimodal Large Language Models (MLLMs) excel at basic Visual Question Answering (VQA), they consistently struggle to grasp the nuanced cultural, emotional, and contextual implications embedded in visual content. This difficulty stems from the task's demand for sophisticated multi-hop reasoning, cultural context, and Theory of Mind (ToM) capabilities, which current models lack. To fill this gap, we propose MetaphorStar, the first end-to-end visual reinforcement learning (RL) framework for image implication tasks. Our framework includes three core components: the fine-grained dataset TFQ-Data, the visual RL method TFQ-GRPO, and the well-structured benchmark TFQ-Bench. Our fully open-source MetaphorStar family, trained using TFQ-GRPO on TFQ-Data, significantly improves performance by an average of 82.6% on the image implication benchmarks. Compared with 20+ mainstream MLLMs, MetaphorStar-32B achieves state-of-the-art (SOTA) on Multiple-Choice Question and Open-Style Question, significantly outperforms the top closed-source model Gemini-3.0-pro on True-False Question. Crucially, our experiments reveal that learning image implication tasks improves the general understanding ability, especially the complex visual reasoning ability. We further provide a systematic analysis of model parameter scaling, training data scaling, and the impact of different model architectures and training strategies, demonstrating the broad applicability of our method. We open-sourced all model weights, datasets, and method code at https://metaphorstar.github.io.</p> + https://arxiv.org/abs/2602.10575 + Wed, 11 Feb 2026 06:59:36 +0000 - VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval - https://arxiv.org/abs/2602.08099 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08099.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Issar Tzachor, Dvir Samuel, Rami Ben-Ari</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Recent studies have adapted generative Multimodal Large Language Models (MLLMs) into embedding extractors for vision tasks, typically through fine-tuning to produce universal representations. However, their performance on video remains inferior to Video Foundation Models (VFMs). In this paper, we focus on leveraging MLLMs for video-text embedding and retrieval. We first conduct a systematic layer-wise analysis, showing that intermediate (pre-trained) MLLM layers already encode substantial task-relevant information. Leveraging this insight, we demonstrate that combining intermediate-layer embeddings with a calibrated MLLM head yields strong zero-shot retrieval performance without any training. Building on these findings, we introduce a lightweight text-based alignment strategy which maps dense video captions to short summaries and enables task-related video-text embedding learning without visual supervision. Remarkably, without any fine-tuning beyond text, our method outperforms current methods, often by a substantial margin, achieving state-of-the-art results across common video retrieval benchmarks.</p> - https://arxiv.org/abs/2602.08099 - Sun, 08 Feb 2026 19:39:32 +0000 + Neural Additive Experts: Context-Gated Experts for Controllable Model Additivity + https://arxiv.org/abs/2602.10585 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10585.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Guangzhi Xiong, Sanchit Sinha, Aidong Zhang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> The trade-off between interpretability and accuracy remains a core challenge in machine learning. Standard Generalized Additive Models (GAMs) offer clear feature attributions but are often constrained by their strictly additive nature, which can limit predictive performance. Introducing feature interactions can boost accuracy yet may obscure individual feature contributions. To address these issues, we propose Neural Additive Experts (NAEs), a novel framework that seamlessly balances interpretability and accuracy. NAEs employ a mixture of experts framework, learning multiple specialized networks per feature, while a dynamic gating mechanism integrates information across features, thereby relaxing rigid additive constraints. Furthermore, we propose targeted regularization techniques to mitigate variance among expert predictions, facilitating a smooth transition from an exclusively additive model to one that captures intricate feature interactions while maintaining clarity in feature attributions. Our theoretical analysis and experiments on synthetic data illustrate the model's flexibility, and extensive evaluations on real-world datasets confirm that NAEs achieve an optimal balance between predictive accuracy and transparent, feature-level explanations. The code is available at https://github.com/Teddy-XiongGZ/NAE.</p> + https://arxiv.org/abs/2602.10585 + Wed, 11 Feb 2026 07:19:25 +0000 - G-LNS: Generative Large Neighborhood Search for LLM-Based Automatic Heuristic Design - https://arxiv.org/abs/2602.08253 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08253.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Baoyun Zhao, He Wang, Liang Zeng</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> While Large Language Models (LLMs) have recently shown promise in Automated Heuristic Design (AHD), existing approaches typically formulate AHD around constructive priority rules or parameterized local search guidance, thereby restricting the search space to fixed heuristic forms. Such designs offer limited capacity for structural exploration, making it difficult to escape deep local optima in complex Combinatorial Optimization Problems (COPs). In this work, we propose G-LNS, a generative evolutionary framework that extends LLM-based AHD to the automated design of Large Neighborhood Search (LNS) operators. Unlike prior methods that evolve heuristics in isolation, G-LNS leverages LLMs to co-evolve tightly coupled pairs of destroy and repair operators. A cooperative evaluation mechanism explicitly captures their interaction, enabling the discovery of complementary operator logic that jointly performs effective structural disruption and reconstruction. Extensive experiments on challenging COP benchmarks, such as Traveling Salesman Problems (TSP) and Capacitated Vehicle Routing Problems (CVRP), demonstrate that G-LNS significantly outperforms LLM-based AHD methods as well as strong classical solvers. The discovered heuristics not only achieve near-optimal solutions with reduced computational budgets but also exhibit robust generalization across diverse and unseen instance distributions.</p> - https://arxiv.org/abs/2602.08253 - Mon, 09 Feb 2026 04:13:35 +0000 + MOSS-Audio-Tokenizer: Scaling Audio Tokenizers for Future Audio Foundation Models + https://arxiv.org/abs/2602.10934 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10934.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yitian Gong, Kuangwei Chen, Zhaoye Fei, Xiaogui Yang, Ke Chen, Yang Wang, Kexin Huang, Mingshu Chen, Ruixiao Li, Qingyuan Cheng, Shimin Li, Xipeng Qiu</p><p><b>Upvotes:</b> 43</p><p><b>Summary:</b> Discrete audio tokenizers are fundamental to empowering large language models with native audio processing and generation capabilities. Despite recent progress, existing approaches often rely on pretrained encoders, semantic distillation, or heterogeneous CNN-based architectures. These designs introduce fixed inductive biases that limit reconstruction fidelity and hinder effective scaling. In this paper, we argue that discrete audio tokenization should be learned fully end-to-end using a homogeneous and scalable architecture. To this end, we first propose CAT (Causal Audio Tokenizer with Transformer), a purely Transformer-based architecture that jointly optimizes the encoder, quantizer, and decoder from scratch for high-fidelity reconstruction. Building on the CAT architecture, we develop MOSS-Audio-Tokenizer, a large-scale audio tokenizer featuring 1.6 billion parameters, pre-trained on 3 million hours of diverse, general audio data. We show that this simple, fully end-to-end approach built from homogeneous, causal Transformer blocks scales gracefully and supports high-fidelity reconstruction across diverse audio domains. Across speech, sound, and music, MOSS-Audio-Tokenizer consistently outperforms prior codecs over a wide range of bitrates, while exhibiting predictable improvements with increased scale. Notably, leveraging the discrete tokens from our model, we develop the first purely autoregressive TTS model that surpasses prior non-autoregressive and cascaded systems. Furthermore, MOSS-Audio-Tokenizer enables competitive ASR performance without auxiliary encoders. Our findings position the CAT architecture as a unified, scalable interface for the next generation of native audio foundation models.</p> + https://arxiv.org/abs/2602.10934 + Wed, 11 Feb 2026 15:13:27 +0000 - Beyond Correctness: Learning Robust Reasoning via Transfer - https://arxiv.org/abs/2602.08489 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08489.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hyunseok Lee, Soheil Abbasloo, Jihoon Tack, Jinwoo Shin</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Reinforcement Learning with Verifiable Rewards (RLVR) has recently strengthened LLM reasoning, but its focus on final answer correctness leaves a critical gap: it does not ensure the robustness of the reasoning process itself. We adopt a simple philosophical view, robust reasoning should remain useful beyond the mind that produced it, and treat reasoning as a form of meaning transfer that must survive truncation, reinterpretation, and continuation. Building on this principle, we introduce Reinforcement Learning with Transferable Reward (RLTR), which operationalizes robustness via transfer reward that tests whether a partial reasoning prefix from one model can guide a separate model to the correct answer. This encourages LLMs to produce reasoning that is stable, interpretable, and genuinely generalizable. Our approach improves sampling consistency while improving final answer accuracy, and it reaches comparable performance in substantially fewer training steps. For example, on MATH500, RLTR achieves a +3.6%p gain in Maj@64 compared to RLVR and matches RLVR's average accuracy with roughly 2.5x fewer training steps, providing both more reliable reasoning and significantly more sample efficient.</p> - https://arxiv.org/abs/2602.08489 - Mon, 09 Feb 2026 10:41:44 +0000 + RISE: Self-Improving Robot Policy with Compositional World Model + https://arxiv.org/abs/2602.11075 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11075.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiazhi Yang, Kunyang Lin, Jinwei Li, Wencong Zhang, Tianwei Lin, Longyan Wu, Zhizhong Su, Hao Zhao, Ya-Qin Zhang, Li Chen, Ping Luo, Xiangyu Yue, Hongyang Li</p><p><b>Upvotes:</b> 18</p><p><b>Summary:</b> Despite the sustained scaling on model capacity and data acquisition, Vision-Language-Action (VLA) models remain brittle in contact-rich and dynamic manipulation tasks, where minor execution deviations can compound into failures. While reinforcement learning (RL) offers a principled path to robustness, on-policy RL in the physical world is constrained by safety risk, hardware cost, and environment reset. To bridge this gap, we present RISE, a scalable framework of robotic reinforcement learning via imagination. At its core is a Compositional World Model that (i) predicts multi-view future via a controllable dynamics model, and (ii) evaluates imagined outcomes with a progress value model, producing informative advantages for the policy improvement. Such compositional design allows state and value to be tailored by best-suited yet distinct architectures and objectives. These components are integrated into a closed-loop self-improving pipeline that continuously generates imaginary rollouts, estimates advantages, and updates the policy in imaginary space without costly physical interaction. Across three challenging real-world tasks, RISE yields significant improvement over prior art, with more than +35% absolute performance increase in dynamic brick sorting, +45% for backpack packing, and +35% for box closing, respectively.</p> + https://arxiv.org/abs/2602.11075 + Wed, 11 Feb 2026 17:43:36 +0000 - TimeChat-Captioner: Scripting Multi-Scene Videos with Time-Aware and Structural Audio-Visual Captions - https://arxiv.org/abs/2602.08711 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08711.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Linli Yao, Yuancheng Wei, Yaojie Zhang, Lei Li, Xinlong Chen, Feifan Song, Ziyue Wang, Kun Ouyang, Yuanxin Liu, Lingpeng Kong, Qi Liu, Pengfei Wan, Kun Gai, Yuanxing Zhang, Xu Sun</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> This paper proposes Omni Dense Captioning, a novel task designed to generate continuous, fine-grained, and structured audio-visual narratives with explicit timestamps. To ensure dense semantic coverage, we introduce a six-dimensional structural schema to create "script-like" captions, enabling readers to vividly imagine the video content scene by scene, akin to a cinematographic screenplay. To facilitate research, we construct OmniDCBench, a high-quality, human-annotated benchmark, and propose SodaM, a unified metric that evaluates time-aware detailed descriptions while mitigating scene boundary ambiguity. Furthermore, we construct a training dataset, TimeChatCap-42K, and present TimeChat-Captioner-7B, a strong baseline trained via SFT and GRPO with task-specific rewards. Extensive experiments demonstrate that TimeChat-Captioner-7B achieves state-of-the-art performance, surpassing Gemini-2.5-Pro, while its generated dense descriptions significantly boost downstream capabilities in audio-visual reasoning (DailyOmni and WorldSense) and temporal grounding (Charades-STA). All datasets, models, and code will be made publicly available at https://github.com/yaolinli/TimeChat-Captioner.</p> - https://arxiv.org/abs/2602.08711 - Mon, 09 Feb 2026 14:21:58 +0000 + Voxtral Realtime + https://arxiv.org/abs/2602.11298 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11298.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Alexander H. Liu, Andy Ehrenberg, Andy Lo, Chen-Yo Sun, Guillaume Lample, Jean-Malo Delignon, Khyathi Raghavi Chandu, Patrick von Platen, Pavankumar Reddy Muddireddy, Rohin Arora, Sanchit Gandhi, Sandeep Subramanian, Soham Ghosh, Srijan Mishra, Abhinav Rastogi, Alan Jeffares, Albert Jiang, Alexandre Sablayrolles, Amélie Héliou, Andrew Bai, Angele Lenglemetz, Anmol Agarwal, Anton Eliseev, Antonia Calvi, Arjun Majumdar, Baptiste Bout, Baptiste Rozière, Baudouin De Monicault, Benjamin Tibi, Clémence Lanfranchi, Connor Chen, Corentin Barreau, Corentin Sautier, Cyprien Courtot, Darius Dabert, Diego de las Casas, Elliot Chane-Sane, Enguerrand Paquin, Faruk Ahmed, Federico Baldassarre, Gabrielle Berrada, Gaëtan Ecrepont, Gauthier Guinet, Genevieve Hayes, Georgii Novikov, Giada Pistilli, Guillaume Martin, Gunjan Dhanuka, Gunshi Gupta, Han Zhou, Indraneel Mukherjee, Irene Zhang, Jaeyoung Kim, Jan Ludziejewski, Jason Rute, Joachim Studnia, John Harvill, Jonas Amar, Josselin Somerville Roberts, Julien Tauran, Karmesh Yadav, Kartik Khandelwal, Kush Jain, Laurence Aitchison, Léonard Blier, Lingxiao Zhao, Louis Martin, Lucile Saulnier, Luyu Gao, Maarten Buyl, Manan Sharma, Margaret Jennings, Marie Pellat, Mark Prins, Mathieu Poirée, Mathilde Guillaumin, Matthieu Dinot, Matthieu Futeral, Maxime Darrin, Maximilian Augustin, Mert Unsal, Mia Chiquier, Nathan Grinsztajn, Neha Gupta, Olivier Bousquet, Olivier Duchenne, Patricia Wang, Paul Jacob, Paul Wambergue, Paula Kurylowicz, Philomène Chagniot, Pierre Stock, Piotr Miłoś, Prateek Gupta, Pravesh Agrawal, Quentin Torroba, Ram Ramrakhya, Rishi Shah, Romain Sauvestre, Roman Soletskyi, Rosalie Millner, Sagar Vaze, Samuel Humeau, Siddharth Gandhi, Sumukh Aithal, Szymon Antoniak, Teven Le Scao, Théo Cachet, Theo Simon Sorg, Thibaut Lavril, Thomas Chabal, Thomas Foubert, Thomas Robert, Thomas Wang, Tim Lawson, Tom Bewley, Tom Edwards, Tyler Wang, Valeriia Nemychnikova, Van Phung, Vedant Nanda, Victor Jouault, Virgile Richard, Vladislav Bataev, Wassim Bouaziz, Wen-Ding Li, William Marshall, Xinghui Li, Xingran Guo, Xinyu Yang, Yannic Neuhaus, Yihan Wang, Zaccharie Ramzi, Zhenlin Xu</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> We introduce Voxtral Realtime, a natively streaming automatic speech recognition model that matches offline transcription quality at sub-second latency. Unlike approaches that adapt offline models through chunking or sliding windows, Voxtral Realtime is trained end-to-end for streaming, with explicit alignment between audio and text streams. Our architecture builds on the Delayed Streams Modeling framework, introducing a new causal audio encoder and Ada RMS-Norm for improved delay conditioning. We scale pretraining to a large-scale dataset spanning 13 languages. At a delay of 480ms, Voxtral Realtime achieves performance on par with Whisper, the most widely deployed offline transcription system. We release the model weights under the Apache 2.0 license.</p> + https://arxiv.org/abs/2602.11298 + Wed, 11 Feb 2026 19:17:10 +0000 - Large Language Lobotomy: Jailbreaking Mixture-of-Experts via Expert Silencing - https://arxiv.org/abs/2602.08741 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08741.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jona te Lintelo, Lichao Wu, Stjepan Picek</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> The rapid adoption of Mixture-of-Experts (MoE) architectures marks a major shift in the deployment of Large Language Models (LLMs). MoE LLMs improve scaling efficiency by activating only a small subset of parameters per token, but their routing structure introduces new safety attack surfaces. We find that safety-critical behaviors in MoE LLMs (e.g., refusal) are concentrated in a small set of experts rather than being uniformly distributed. Building on this, we propose Large Language Lobotomy (L^3), a training-free, architecture-agnostic attack that compromises safety alignment by exploiting expert routing dynamics. L^3 learns routing patterns that correlate with refusal, attributes safety behavior to specific experts, and adaptively silences the most safety-relevant experts until harmful outputs are produced. We evaluate L^3 on eight state-of-the-art open-source MoE LLMs and show that our adaptive expert silencing increases average attack success from 7.3% to 70.4%, reaching up to 86.3%, outperforming prior training-free MoE jailbreak methods. Moreover, bypassing guardrails typically requires silencing fewer than 20% of layer-wise experts while largely preserving general language utility. These results reveal a fundamental tension between efficiency-driven MoE design and robust safety alignment and motivate distributing safety mechanisms more robustly in future MoE LLMs with architecture- and routing-aware methods.</p> - https://arxiv.org/abs/2602.08741 - Mon, 09 Feb 2026 14:42:11 +0000 + MolmoSpaces: A Large-Scale Open Ecosystem for Robot Navigation and Manipulation + https://arxiv.org/abs/2602.11337 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11337.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yejin Kim, Wilbert Pumacay, Omar Rayyan, Max Argus, Winson Han, Eli VanderBilt, Jordi Salvador, Abhay Deshpande, Rose Hendrix, Snehal Jauhri, Shuo Liu, Nur Muhammad Mahi Shafiullah, Maya Guru, Ainaz Eftekhar, Karen Farley, Donovan Clay, Jiafei Duan, Arjun Guru, Piper Wolters, Alvaro Herrasti, Ying-Chun Lee, Georgia Chalvatzaki, Yuchen Cui, Ali Farhadi, Dieter Fox, Ranjay Krishna</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Deploying robots at scale demands robustness to the long tail of everyday situations. The countless variations in scene layout, object geometry, and task specifications that characterize real environments are vast and underrepresented in existing robot benchmarks. Measuring this level of generalization requires infrastructure at a scale and diversity that physical evaluation alone cannot provide. We introduce MolmoSpaces, a fully open ecosystem to support large-scale benchmarking of robot policies. MolmoSpaces consists of over 230k diverse indoor environments, ranging from handcrafted household scenes to procedurally generated multiroom houses, populated with 130k richly annotated object assets, including 48k manipulable objects with 42M stable grasps. Crucially, these environments are simulator-agnostic, supporting popular options such as MuJoCo, Isaac, and ManiSkill. The ecosystem supports the full spectrum of embodied tasks: static and mobile manipulation, navigation, and multiroom long-horizon tasks requiring coordinated perception, planning, and interaction across entire indoor environments. We also design MolmoSpaces-Bench, a benchmark suite of 8 tasks in which robots interact with our diverse scenes and richly annotated objects. Our experiments show MolmoSpaces-Bench exhibits strong sim-to-real correlation (R = 0.96, ho = 0.98), confirm newer and stronger zero-shot policies outperform earlier versions in our benchmarks, and identify key sensitivities to prompt phrasing, initial joint positions, and camera occlusion. Through MolmoSpaces and its open-source assets and tooling, we provide a foundation for scalable data generation, policy training, and benchmark creation for robot learning research.</p> + https://arxiv.org/abs/2602.11337 + Wed, 11 Feb 2026 20:16:31 +0000 - StealthRL: Reinforcement Learning Paraphrase Attacks for Multi-Detector Evasion of AI-Text Detectors - https://arxiv.org/abs/2602.08934 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08934.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Suraj Ranganath, Atharv Ramesh</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> AI-text detectors face a critical robustness challenge: adversarial paraphrasing attacks that preserve semantics while evading detection. We introduce StealthRL, a reinforcement learning framework that stress-tests detector robustness under realistic adversarial conditions. StealthRL trains a paraphrase policy against a multi-detector ensemble using Group Relative Policy Optimization (GRPO) with LoRA adapters on Qwen3-4B, optimizing a composite reward that balances detector evasion with semantic preservation. We evaluate six attack settings (M0-M5) against three detector families (RoBERTa, FastDetectGPT, and Binoculars) at the security-relevant 1% false positive rate operating point. StealthRL achieves near-zero detection (0.001 mean TPR@1%FPR), reduces mean AUROC from 0.74 to 0.27, and attains a 99.9% attack success rate. Critically, attacks transfer to a held-out detector family not seen during training, revealing shared architectural vulnerabilities rather than detector-specific brittleness. We additionally conduct LLM-based quality evaluation via Likert scoring, analyze detector score distributions to explain why evasion succeeds, and provide per-detector AUROC with bootstrap confidence intervals. Our results expose significant robustness gaps in current AI-text detection and establish StealthRL as a principled adversarial evaluation protocol. Code and evaluation pipeline are publicly available at https://github.com/suraj-ranganath/StealthRL.</p> - https://arxiv.org/abs/2602.08934 - Mon, 09 Feb 2026 17:33:46 +0000 + Multimodal Fact-Level Attribution for Verifiable Reasoning + https://arxiv.org/abs/2602.11509 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11509.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> David Wan, Han Wang, Ziyang Wang, Elias Stengel-Eskin, Hyunji Lee, Mohit Bansal</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Multimodal large language models (MLLMs) are increasingly used for real-world tasks involving multi-step reasoning and long-form generation, where reliability requires grounding model outputs in heterogeneous input sources and verifying individual factual claims. However, existing multimodal grounding benchmarks and evaluation methods focus on simplified, observation-based scenarios or limited modalities and fail to assess attribution in complex multimodal reasoning. We introduce MuRGAt (Multimodal Reasoning with Grounded Attribution), a benchmark for evaluating fact-level multimodal attribution in settings that require reasoning beyond direct observation. Given inputs spanning video, audio, and other modalities, MuRGAt requires models to generate answers with explicit reasoning and precise citations, where each citation specifies both modality and temporal segments. To enable reliable assessment, we introduce an automatic evaluation framework that strongly correlates with human judgments. Benchmarking with human and automated scores reveals that even strong MLLMs frequently hallucinate citations despite correct reasoning. Moreover, we observe a key trade-off: increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting a significant gap between internal reasoning and verifiable attribution.</p> + https://arxiv.org/abs/2602.11509 + Thu, 12 Feb 2026 03:10:02 +0000 - When Actions Go Off-Task: Detecting and Correcting Misaligned Actions in Computer-Use Agents - https://arxiv.org/abs/2602.08995 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.08995.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuting Ning, Jaylen Jones, Zhehao Zhang, Chentao Ye, Weitong Ruan, Junyi Li, Rahul Gupta, Huan Sun</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Computer-use agents (CUAs) have made tremendous progress in the past year, yet they still frequently produce misaligned actions that deviate from the user's original intent. Such misaligned actions may arise from external attacks (e.g., indirect prompt injection) or from internal limitations (e.g., erroneous reasoning). They not only expose CUAs to safety risks, but also degrade task efficiency and reliability. This work makes the first effort to define and study misaligned action detection in CUAs, with comprehensive coverage of both externally induced and internally arising misaligned actions. We further identify three common categories in real-world CUA deployment and construct MisActBench, a benchmark of realistic trajectories with human-annotated, action-level alignment labels. Moreover, we propose DeAction, a practical and universal guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. DeAction outperforms all existing baselines across offline and online evaluations with moderate latency overhead: (1) On MisActBench, it outperforms baselines by over 15% absolute in F1 score; (2) In online evaluation, it reduces attack success rate by over 90% under adversarial settings while preserving or even improving task success rate in benign environments.</p> - https://arxiv.org/abs/2602.08995 - Mon, 09 Feb 2026 18:41:15 +0000 + Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use + https://arxiv.org/abs/2602.11541 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11541.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hanbing Liu, Chunhao Tian, Nan An, Ziyuan Wang, Pinyan Lu, Changyuan Yu, Qi Qi</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> We study budget-constrained tool-augmented agents, where a large language model must solve multi-step tasks by invoking external tools under a strict monetary budget. We formalize this setting as sequential decision making in context space with priced and stochastic tool executions, making direct planning intractable due to massive state-action spaces, high variance of outcomes and prohibitive exploration cost. To address these challenges, we propose INTENT, an inference-time planning framework that leverages an intention-aware hierarchical world model to anticipate future tool usage, risk-calibrated cost, and guide decisions online. Across cost-augmented StableToolBench, INTENT strictly enforces hard budget feasibility while substantially improving task success over baselines, and remains robust under dynamic market shifts such as tool price changes and varying budgets.</p> + https://arxiv.org/abs/2602.11541 + Thu, 12 Feb 2026 04:01:30 +0000 - ArcFlow: Unleashing 2-Step Text-to-Image Generation via High-Precision Non-Linear Flow Distillation - https://arxiv.org/abs/2602.09014 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09014.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zihan Yang, Shuyuan Tu, Licheng Zhang, Qi Dai, Yu-Gang Jiang, Zuxuan Wu</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Diffusion models have achieved remarkable generation quality, but they suffer from significant inference cost due to their reliance on multiple sequential denoising steps, motivating recent efforts to distill this inference process into a few-step regime. However, existing distillation methods typically approximate the teacher trajectory by using linear shortcuts, which makes it difficult to match its constantly changing tangent directions as velocities evolve across timesteps, thereby leading to quality degradation. To address this limitation, we propose ArcFlow, a few-step distillation framework that explicitly employs non-linear flow trajectories to approximate pre-trained teacher trajectories. Concretely, ArcFlow parameterizes the velocity field underlying the inference trajectory as a mixture of continuous momentum processes. This enables ArcFlow to capture velocity evolution and extrapolate coherent velocities to form a continuous non-linear trajectory within each denoising step. Importantly, this parameterization admits an analytical integration of this non-linear trajectory, which circumvents numerical discretization errors and results in high-precision approximation of the teacher trajectory. To train this parameterization into a few-step generator, we implement ArcFlow via trajectory distillation on pre-trained teacher models using lightweight adapters. This strategy ensures fast, stable convergence while preserving generative diversity and quality. Built on large-scale models (Qwen-Image-20B and FLUX.1-dev), ArcFlow only fine-tunes on less than 5% of original parameters and achieves a 40x speedup with 2 NFEs over the original multi-step teachers without significant quality degradation. Experiments on benchmarks show the effectiveness of ArcFlow both qualitatively and quantitatively.</p> - https://arxiv.org/abs/2602.09014 - Mon, 09 Feb 2026 18:56:14 +0000 + Pretraining A Large Language Model using Distributed GPUs: A Memory-Efficient Decentralized Paradigm + https://arxiv.org/abs/2602.11543 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11543.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jinrui Zhang, Chaodong Xiao, Aoqi Wu, Xindong Zhang, Lei Zhang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Pretraining large language models (LLMs) typically requires centralized clusters with thousands of high-memory GPUs (e.g., H100/A100). Recent decentralized training methods reduce communication overhead by employing federated optimization; however, they still need to train the entire model on each node, remaining constrained by GPU memory limitations. In this work, we propose SParse Expert Synchronization (SPES), a memory-efficient decentralized framework for pretraining mixture-of-experts (MoE) LLMs. SPES trains only a subset of experts per node, substantially lowering the memory footprint. Each node updates its local experts and periodically synchronizes with other nodes, eliminating full-parameter transmission while ensuring efficient knowledge sharing. To accelerate convergence, we introduce an expert-merging warm-up strategy, where experts exchange knowledge early in training, to rapidly establish foundational capabilities. With SPES, we train a 2B-parameter MoE LLM using 16 standalone 48GB GPUs over internet connections, which achieves competitive performance with centrally trained LLMs under similar computational budgets. We further demonstrate scalability by training a 7B model from scratch and a 9B model upcycled from a dense checkpoint, both of which match prior centralized baselines. Our code is available at https://github.com/zjr2000/SPES.</p> + https://arxiv.org/abs/2602.11543 + Thu, 12 Feb 2026 04:02:45 +0000 - EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies - https://arxiv.org/abs/2602.09514 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09514.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xavier Hu, Jinxiang Xia, Shengze Xu, Kangqi Song, Yishuo Yuan, Guibin Zhang, JinCheng Ren, Boyu Feng, Li Lu, Tieyong Zeng, Jiaheng Liu, Minghao Liu, He Zhu, Yuchen Eleanor Jiang, Wei Wang, Wangchunshu Zhou</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Long-horizon planning is widely recognized as a core capability of autonomous LLM-based agents; however, current evaluation frameworks suffer from being largely episodic, domain-specific, or insufficiently grounded in persistent economic dynamics. We introduce EcoGym, a generalizable benchmark for continuous plan-and-execute decision making in interactive economies. EcoGym comprises three diverse environments: Vending, Freelance, and Operation, implemented in a unified decision-making process with standardized interfaces, and budgeted actions over an effectively unbounded horizon (1000+ steps if 365 day-loops for evaluation). The evaluation of EcoGym is based on business-relevant outcomes (e.g., net worth, income, and DAU), targeting long-term strategic coherence and robustness under partial observability and stochasticity. Experiments across eleven leading LLMs expose a systematic tension: no single model dominates across all three scenarios. Critically, we find that models exhibit significant suboptimality in either high-level strategies or efficient actions executions. EcoGym is released as an open, extensible testbed for transparent long-horizon agent evaluation and for studying controllability-utility trade-offs in realistic economic settings.</p> - https://arxiv.org/abs/2602.09514 - Tue, 10 Feb 2026 08:12:23 +0000 + ABot-N0: Technical Report on the VLA Foundation Model for Versatile Embodied Navigation + https://arxiv.org/abs/2602.11598 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11598.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zedong Chu, Shichao Xie, Xiaolong Wu, Yanfen Shen, Minghua Luo, Zhengbo Wang, Fei Liu, Xiaoxu Leng, Junjun Hu, Mingyang Yin, Jia Lu, Yingnan Guo, Kai Yang, Jiawei Han, Xu Chen, Yanqing Zhu, Yuxiang Zhao, Xin Liu, Yirong Yang, Ye He, Jiahang Wang, Yang Cai, Tianlin Zhang, Li Gao, Liu Liu, Mingchao Sun, Fan Jiang, Chiyu Wang, Zhicheng Liu, Hongyu Pan, Honglin Han, Zhining Gu, Kuan Yang, Jianfang Zhang, Di Jing, Zihao Guan, Wei Guo, Guoqing Liu, Di Yang, Xiangpo Yang, Menglin Yang, Hongguang Xing, Weiguo Li, Mu Xu</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Embodied navigation has long been fragmented by task-specific architectures. We introduce ABot-N0, a unified Vision-Language-Action (VLA) foundation model that achieves a ``Grand Unification'' across 5 core tasks: Point-Goal, Object-Goal, Instruction-Following, POI-Goal, and Person-Following. ABot-N0 utilizes a hierarchical ``Brain-Action'' architecture, pairing an LLM-based Cognitive Brain for semantic reasoning with a Flow Matching-based Action Expert for precise, continuous trajectory generation. To support large-scale learning, we developed the ABot-N0 Data Engine, curating 16.9M expert trajectories and 5.0M reasoning samples across 7,802 high-fidelity 3D scenes (10.7 km^2). ABot-N0 achieves new SOTA performance across 7 benchmarks, significantly outperforming specialized models. Furthermore, our Agentic Navigation System integrates a planner with hierarchical topological memory, enabling robust, long-horizon missions in dynamic real-world environments.</p> + https://arxiv.org/abs/2602.11598 + Thu, 12 Feb 2026 05:30:20 +0000 - Stroke3D: Lifting 2D strokes into rigged 3D model via latent diffusion models - https://arxiv.org/abs/2602.09713 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09713.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ruisi Zhao, Haoren Zheng, Zongxin Yang, Hehe Fan, Yi Yang</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Rigged 3D assets are fundamental to 3D deformation and animation. However, existing 3D generation methods face challenges in generating animatable geometry, while rigging techniques lack fine-grained structural control over skeleton creation. To address these limitations, we introduce Stroke3D, a novel framework that directly generates rigged meshes from user inputs: 2D drawn strokes and a descriptive text prompt. Our approach pioneers a two-stage pipeline that separates the generation into: 1) Controllable Skeleton Generation, we employ the Skeletal Graph VAE (Sk-VAE) to encode the skeleton's graph structure into a latent space, where the Skeletal Graph DiT (Sk-DiT) generates a skeletal embedding. The generation process is conditioned on both the text for semantics and the 2D strokes for explicit structural control, with the VAE's decoder reconstructing the final high-quality 3D skeleton; and 2) Enhanced Mesh Synthesis via TextuRig and SKA-DPO, where we then synthesize a textured mesh conditioned on the generated skeleton. For this stage, we first enhance an existing skeleton-to-mesh model by augmenting its training data with TextuRig: a dataset of textured and rigged meshes with captions, curated from Objaverse-XL. Additionally, we employ a preference optimization strategy, SKA-DPO, guided by a skeleton-mesh alignment score, to further improve geometric fidelity. Together, our framework enables a more intuitive workflow for creating ready to animate 3D content. To the best of our knowledge, our work is the first to generate rigged 3D meshes conditioned on user-drawn 2D strokes. Extensive experiments demonstrate that Stroke3D produces plausible skeletons and high-quality meshes.</p> - https://arxiv.org/abs/2602.09713 - Tue, 10 Feb 2026 12:17:00 +0000 + ScalSelect: Scalable Training-Free Multimodal Data Selection for Efficient Visual Instruction Tuning + https://arxiv.org/abs/2602.11636 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11636.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Changti Wu, Jiahuai Mao, Yuzhuo Miao, Shijie Lian, Bin Yu, Xiaopeng Lin, Cong Huang, Lei Zhang, Kai Chen</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Large-scale Visual Instruction Tuning (VIT) has become a key paradigm for advancing the performance of vision-language models (VLMs) across various multimodal tasks. However, training on the large-scale datasets is computationally expensive and inefficient due to redundancy in the data, which motivates the need for multimodal data selection to improve training efficiency. Existing data selection methods for VIT either require costly training or gradient computation. Training-free alternatives often depend on proxy models or datasets, instruction-agnostic representations, and pairwise similarity with quadratic complexity, limiting scalability and representation fidelity. In this work, we propose ScalSelect, a scalable training-free multimodal data selection method with linear-time complexity with respect to the number of samples, eliminating the need for external models or auxiliary datasets. ScalSelect first constructs sample representations by extracting visual features most attended by instruction tokens in the target VLM, capturing instruction-relevant information. It then identifies samples whose representations best approximate the dominant subspace of the full dataset representations, enabling scalable importance scoring without pairwise comparisons. Extensive experiments across multiple VLMs, datasets, and selection budgets demonstrate that ScalSelect achieves over 97.5% of the performance of training on the full dataset using only 16% of the data, and even outperforms full-data training in some settings. The code is available at https://github.com/ChangtiWu/ScalSelect{ScalSelect}.</p> + https://arxiv.org/abs/2602.11636 + Thu, 12 Feb 2026 06:38:49 +0000 - QP-OneModel: A Unified Generative LLM for Multi-Task Query Understanding in Xiaohongshu Search - https://arxiv.org/abs/2602.09901 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.09901.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jianzhao Huang, Xiaorui Huang, Fei Zhao, Yunpeng Liu, Hui Zhang, Fangcheng Shi, Congfeng Li, Zechen Sun, Yi Wu, Yao Hu, Yunhan Bai, Shaosheng Cao</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Query Processing (QP) bridges user intent and content supply in large-scale Social Network Service (SNS) search engines. Traditional QP systems rely on pipelines of isolated discriminative models (e.g., BERT), suffering from limited semantic understanding and high maintenance overhead. While Large Language Models (LLMs) offer a potential solution, existing approaches often optimize sub-tasks in isolation, neglecting intrinsic semantic synergy and necessitating independent iterations. Moreover, standard generative methods often lack grounding in SNS scenarios, failing to bridge the gap between open-domain corpora and informal SNS linguistic patterns, while struggling to adhere to rigorous business definitions. We present QP-OneModel, a Unified Generative LLM for Multi-Task Query Understanding in the SNS domain. We reformulate heterogeneous sub-tasks into a unified sequence generation paradigm, adopting a progressive three-stage alignment strategy culminating in multi-reward Reinforcement Learning. Furthermore, QP-OneModel generates intent descriptions as a novel high-fidelity semantic signal, effectively augmenting downstream tasks such as query rewriting and ranking. Offline evaluations show QP-OneModel achieves a 7.35% overall gain over discriminative baselines, with significant F1 boosts in NER (+9.01%) and Term Weighting (+9.31%). It also exhibits superior generalization, surpassing a 32B model by 7.60% accuracy on unseen tasks. Fully deployed at Xiaohongshu, online A/B tests confirm its industrial value, optimizing retrieval relevance (DCG) by 0.21% and lifting user retention by 0.044%.</p> - https://arxiv.org/abs/2602.09901 - Tue, 10 Feb 2026 15:38:17 +0000 + ThinkRouter: Efficient Reasoning via Routing Thinking between Latent and Discrete Spaces + https://arxiv.org/abs/2602.11683 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11683.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xin Xu, Tong Yu, Xiang Chen, Haoliang Wang, Julian McAuley, Saayan Mitra</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Recent work explores latent reasoning to improve reasoning efficiency by replacing explicit reasoning trajectories with continuous representations in a latent space, yet its effectiveness varies across settings. Analysis of model confidence dynamics under latent reasoning reveals that thinking trajectories ending in incorrect answers contain fewer low-confidence steps than those ending in correct answers. Meanwhile, we suggest that soft embeddings aggregated by multiple low-confidence thinking alternatives may introduce and propagate noise, leading to high confidence in unreliable reasoning trajectories. Motivated by these observations, ThinkRouter, an inference-time confidence-aware routing mechanism is proposed to avoid high confidence and noise for efficient reasoning. ThinkRouter routes thinking to the discrete token space when model confidence is low, and to the latent space otherwise. Extensive experiments on STEM reasoning and coding benchmarks across diverse large reasoning models demonstrate that ThinkRouter outperforms explicit CoT, random routing, and latent reasoning baselines in terms of accuracy, achieving an average improvement of 19.70 points in Pass@1, while reducing generation length by up to 15.55%. Further comprehensive analysis reveals that ThinkRouter can calibrate errors arising from explicit CoT and latent reasoning, and accelerates end-of-thinking token generation by globally lowering model confidence.</p> + https://arxiv.org/abs/2602.11683 + Thu, 12 Feb 2026 08:01:01 +0000 - Towards Autonomous Mathematics Research - https://arxiv.org/abs/2602.10177 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10177.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tony Feng, Trieu H. Trinh, Garrett Bingham, Dawsen Hwang, Yuri Chervonyi, Junehyuk Jung, Joonkyung Lee, Carlo Pagano, Sang-hyun Kim, Federico Pasqualotto, Sergei Gukov, Jonathan N. Lee, Junsu Kim, Kaiying Hou, Golnaz Ghiasi, Yi Tay, YaGuang Li, Chenkai Kuang, Yuan Liu, Hanzhao, Lin, Evan Zheran Liu, Nigamaa Nayakanti, Xiaomeng Yang, Heng-tze Cheng, Demis Hassabis, Koray Kavukcuoglu, Quoc V. Le, Thang Luong</p><p><b>Upvotes:</b> 24</p><p><b>Summary:</b> Recent advances in foundational models have yielded reasoning systems capable of achieving a gold-medal standard at the International Mathematical Olympiad. The transition from competition-level problem-solving to professional research, however, requires navigating vast literature and constructing long-horizon proofs. In this work, we introduce Aletheia, a math research agent that iteratively generates, verifies, and revises solutions end-to-end in natural language. Specifically, Aletheia is powered by an advanced version of Gemini Deep Think for challenging reasoning problems, a novel inference-time scaling law that extends beyond Olympiad-level problems, and intensive tool use to navigate the complexities of mathematical research. We demonstrate the capability of Aletheia from Olympiad problems to PhD-level exercises and most notably, through several distinct milestones in AI-assisted mathematics research: (a) a research paper (Feng26) generated by AI without any human intervention in calculating certain structure constants in arithmetic geometry called eigenweights; (b) a research paper (LeeSeo26) demonstrating human-AI collaboration in proving bounds on systems of interacting particles called independent sets; and (c) an extensive semi-autonomous evaluation (Feng et al., 2026a) of 700 open problems on Bloom's Erdos Conjectures database, including autonomous solutions to four open questions. In order to help the public better understand the developments pertaining to AI and mathematics, we suggest codifying standard levels quantifying autonomy and novelty of AI-assisted results. We conclude with reflections on human-AI collaboration in mathematics.</p> - https://arxiv.org/abs/2602.10177 - Tue, 10 Feb 2026 18:50:15 +0000 + Thinking with Drafting: Optical Decompression via Logical Reconstruction + https://arxiv.org/abs/2602.11731 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11731.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jingxuan Wei, Honghao He, Caijun Jia, Siyuan Li, Zheng Sun, Yuhang Xu, Yuanyuan Lin, Linzhuang Sun, Yuchen Wu, Bihui Yu, Xiangxiang Zhang, Cheng Tan</p><p><b>Upvotes:</b> 27</p><p><b>Summary:</b> Existing multimodal large language models have achieved high-fidelity visual perception and exploratory visual generation. However, a precision paradox persists in complex reasoning tasks: optical perception systems transcribe symbols without capturing logical topology, while pixel-based generative models produce visual artifacts lacking mathematical exactness. To bridge this gap, we propose that reasoning over visual inputs be reconceptualized as optical decompression-the process of reconstructing latent logical structures from compressed visual tokens. Guided by the axiom that Parsing is Reasoning, we introduce Thinking with Drafting (TwD), which utilizes a minimalist Domain-Specific Language (DSL) as a grounding intermediate representation. Unlike standard approaches that hallucinate answers directly, TwD forces the model to draft its mental model into executable code, rendering deterministic visual proofs for self-verification. To validate this, we present VisAlg, a visual algebra benchmark. Experiments demonstrate that TwD serve as a superior cognitive scaffold. Our work establishes a closed-loop system where visual generation acts not as a creative output but as a logical verifier, offering a generalizable path for visual reasoning.</p> + https://arxiv.org/abs/2602.11731 + Thu, 12 Feb 2026 08:54:02 +0000 - When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models - https://arxiv.org/abs/2602.10179 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10179.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiacheng Hou, Yining Sun, Ruochong Jin, Haochen Han, Fangming Liu, Wai Kin Victor Chan, Alex Jinpeng Wang</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Recent advances in large image editing models have shifted the paradigm from text-driven instructions to vision-prompt editing, where user intent is inferred directly from visual inputs such as marks, arrows, and visual-text prompts. While this paradigm greatly expands usability, it also introduces a critical and underexplored safety risk: the attack surface itself becomes visual. In this work, we propose Vision-Centric Jailbreak Attack (VJA), the first visual-to-visual jailbreak attack that conveys malicious instructions purely through visual inputs. To systematically study this emerging threat, we introduce IESBench, a safety-oriented benchmark for image editing models. Extensive experiments on IESBench demonstrate that VJA effectively compromises state-of-the-art commercial models, achieving attack success rates of up to 80.9% on Nano Banana Pro and 70.1% on GPT-Image-1.5. To mitigate this vulnerability, we propose a training-free defense based on introspective multimodal reasoning, which substantially improves the safety of poorly aligned models to a level comparable with commercial systems, without auxiliary guard models and with negligible computational overhead. Our findings expose new vulnerabilities, provide both a benchmark and practical defense to advance safe and trustworthy modern image editing systems. Warning: This paper contains offensive images created by large image editing models.</p> - https://arxiv.org/abs/2602.10179 - Tue, 10 Feb 2026 18:59:55 +0000 + Adapting Vision-Language Models for E-commerce Understanding at Scale + https://arxiv.org/abs/2602.11733 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11733.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Matteo Nulli, Vladimir Orshulevich, Tala Bazazo, Christian Herold, Michael Kozielski, Marcin Mazur, Szymon Tuzel, Cees G. M. Snoek, Seyyed Hadi Hashemi, Omar Javed, Yannick Versley, Shahram Khadivi</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> E-commerce product understanding demands by nature, strong multimodal comprehension from text, images, and structured attributes. General-purpose Vision-Language Models (VLMs) enable generalizable multimodal latent modelling, yet there is no documented, well-known strategy for adapting them to the attribute-centric, multi-image, and noisy nature of e-commerce data, without sacrificing general performance. In this work, we show through a large-scale experimental study, how targeted adaptation of general VLMs can substantially improve e-commerce performance while preserving broad multimodal capabilities. Furthermore, we propose a novel extensive evaluation suite covering deep product understanding, strict instruction following, and dynamic attribute extraction.</p> + https://arxiv.org/abs/2602.11733 + Thu, 12 Feb 2026 08:59:22 +0000 - Internalizing Meta-Experience into Memory for Guided Reinforcement Learning in Large Language Models - https://arxiv.org/abs/2602.10224 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10224.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shiting Huang, Zecheng Li, Yu Zeng, Qingnan Ren, Zhen Fang, Qisheng Su, Kou Shi, Lin Chen, Zehui Chen, Feng Zhao</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective approach for enhancing the reasoning capabilities of Large Language Models (LLMs). Despite its efficacy, RLVR faces a meta-learning bottleneck: it lacks mechanisms for error attribution and experience internalization intrinsic to the human learning cycle beyond practice and verification, thereby limiting fine-grained credit assignment and reusable knowledge formation. We term such reusable knowledge representations derived from past errors as meta-experience. Based on this insight, we propose Meta-Experience Learning (MEL), a novel framework that incorporates self-distilled meta-experience into the model's parametric memory. Building upon standard RLVR, we introduce an additional design that leverages the LLM's self-verification capability to conduct contrastive analysis on paired correct and incorrect trajectories, identify the precise bifurcation points where reasoning errors arise, and summarize them into generalizable meta-experience. The meta-experience is further internalized into the LLM's parametric memory by minimizing the negative log-likelihood, which induces a language-modeled reward signal that bridges correct and incorrect reasoning trajectories and facilitates effective knowledge reuse. Experimental results demonstrate that MEL achieves consistent improvements on benchmarks, yielding 3.92%--4.73% Pass@1 gains across varying model sizes.</p> - https://arxiv.org/abs/2602.10224 - Tue, 10 Feb 2026 19:16:09 +0000 + Think Longer to Explore Deeper: Learn to Explore In-Context via Length-Incentivized Reinforcement Learning + https://arxiv.org/abs/2602.11748 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11748.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Futing Wang, Jianhao Yan, Yun Luo, Ganqu Cui, Zhi Wang, Xiaoye Qu, Yue Zhang, Yu Cheng, Tao Lin</p><p><b>Upvotes:</b> 24</p><p><b>Summary:</b> Achieving effective test-time scaling requires models to engage in In-Context Exploration -- the intrinsic ability to generate, verify, and refine multiple reasoning hypotheses within a single continuous context. Grounded in State Coverage theory, our analysis identifies a critical bottleneck to enabling this capability: while broader state coverage requires longer reasoning trajectories, the probability of sampling such sequences decays exponentially during autoregressive generation, a phenomenon we term the ``Shallow Exploration Trap''. To bridge this gap, we propose Length-Incentivized Exploration(\method). This simple yet effective recipe explicitly encourages models to explore more via a length-based reward coupled with a redundancy penalty, thereby maximizing state coverage in two-step manner. Comprehensive experiments across different models (Qwen3, Llama) demonstrate that \method effectively incentivize in-context exploration. As a result, our method achieves an average improvement of 4.4\% on in-domain tasks and a 2.7\% gain on out-of-domain benchmarks.</p> + https://arxiv.org/abs/2602.11748 + Thu, 12 Feb 2026 09:24:32 +0000 - Latent Thoughts Tuning: Bridging Context and Reasoning with Fused Information in Latent Tokens - https://arxiv.org/abs/2602.10229 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10229.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Weihao Liu, Dehai Min, Lu Cheng</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> While explicit Chain-of-Thought (CoT) equips Large Language Models (LLMs) with strong reasoning capabilities, it requires models to verbalize every intermediate step in text tokens, constraining the model thoughts to the discrete vocabulary space. Recently, reasoning in continuous latent space has emerged as a promising alternative, enabling more robust inference and flexible computation beyond discrete token constraints. However, current latent paradigms often suffer from feature collapse and instability, stemming from distribution mismatches when recurrently using hidden states as the input embeddings, or alignment issues when relying on assistant models. To address this, we propose Latent Thoughts Tuning (LT-Tuning), a framework that redefines how latent thoughts are constructed and deployed. Instead of relying solely on raw hidden states, our method introduces a Context-Prediction-Fusion mechanism that jointly leveraging contextual hidden states and predictive semantic guidance from the vocabulary embedding space. Combined with a progressive three-stage curriculum learning pipeline, LT-Tuning also enables dynamically switching between latent and explicit thinking modes. Experiments demonstrate that our method outperforms existing latent reasoning baselines, effectively mitigating feature collapse and achieving robust reasoning accuracy.</p> - https://arxiv.org/abs/2602.10229 - Tue, 10 Feb 2026 19:19:10 +0000 + MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling + https://arxiv.org/abs/2602.11761 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11761.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> MiniCPM Team, Wenhao An, Yingfa Chen, Yewei Fang, Jiayi Li, Xin Li, Yaohui Li, Yishan Li, Yuxuan Li, Biyuan Lin, Chuan Liu, Hezi Liu, Siyuan Liu, Hongya Lyu, Yinxu Pan, Shixin Ren, Xingyu Shen, Zhou Su, Haojun Sun, Yangang Sun, Zhen Leng Thai, Xin Tian, Rui Wang, Xiaorong Wang, Yudong Wang, Bo Wu, Xiaoyue Xu, Dong Xu, Shuaikang Xue, Jiawei Yang, Bowen Zhang, Jinqian Zhang, Letian Zhang, Shengnan Zhang, Xinyu Zhang, Xinyuan Zhang, Zhu Zhang, Hengyu Zhao, Jiacheng Zhao, Jie Zhou, Zihan Zhou, Shuo Wang, Chaojun Xiao, Xu Han, Zhiyuan Liu, Maosong Sun</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.</p> + https://arxiv.org/abs/2602.11761 + Thu, 12 Feb 2026 09:37:05 +0000 - Blockwise Advantage Estimation for Multi-Objective RL with Verifiable Rewards - https://arxiv.org/abs/2602.10231 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10231.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kirill Pavlenko, Alexander Golubev, Simon Karasik, Boris Yangel</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Group Relative Policy Optimization (GRPO) assigns a single scalar advantage to all tokens in a completion. For structured generations with explicit segments and objectives, this couples unrelated reward signals across segments, leading to objective interference and misattributed credit. We propose Blockwise Advantage Estimation, a family of GRPO-compatible methods that assigns each objective its own advantage and applies it only to the tokens in the corresponding text block, reducing reliance on hand-designed scalar rewards and scaling naturally to additional objectives. A key challenge is estimating advantages for later blocks whose rewards are conditioned on sampled prefixes; standard unbiased approaches require expensive nested rollouts from intermediate states. Concretely, we introduce an Outcome-Conditioned Baseline that approximates intermediate state values using only within-group statistics by stratifying samples according to a prefix-derived intermediate outcome. On math tasks with uncertainty estimation, our method mitigates reward interference, is competitive with a state-of-the-art reward-designed approach, and preserves test-time gains from confidence-weighted ensembling. More broadly, it provides a modular recipe for optimizing sequential objectives in structured generations without additional rollouts.</p> - https://arxiv.org/abs/2602.10231 - Tue, 10 Feb 2026 19:22:37 +0000 + Detecting RLVR Training Data via Structural Convergence of Reasoning + https://arxiv.org/abs/2602.11792 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11792.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hongbo Zhang, Yue Yang, Jianhao Yan, Guangsheng Bao, Yue Zhang, Yue Zhang</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Reinforcement learning with verifiable rewards (RLVR) is central to training modern reasoning models, but the undisclosed training data raises concerns about benchmark contamination. Unlike pretraining methods, which optimize models using token-level probabilities, RLVR fine-tunes models based on reward feedback from self-generated reasoning trajectories, making conventional likelihood-based detection methods less effective. We show that RLVR induces a distinctive behavioral signature: prompts encountered during RLVR training result in more rigid and similar generations, while unseen prompts retain greater diversity. We introduce Min-kNN Distance, a simple black-box detector that quantifies this collapse by sampling multiple completions for a given prompt and computing the average of the k smallest nearest-neighbor edit distances. Min-kNN Distance requires no access to the reference model or token probabilities. Experiments across multiple RLVR-trained reasoning models show that Min-kNN Distance reliably distinguishes RL-seen examples from unseen ones and outperforms existing membership inference and RL contamination detection baselines.</p> + https://arxiv.org/abs/2602.11792 + Thu, 12 Feb 2026 10:17:32 +0000 - LiveMedBench: A Contamination-Free Medical Benchmark for LLMs with Automated Rubric Evaluation - https://arxiv.org/abs/2602.10367 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10367.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhiling Yan, Dingjie Song, Zhe Fang, Yisheng Ji, Xiang Li, Quanzheng Li, Lichao Sun</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> The deployment of Large Language Models (LLMs) in high-stakes clinical settings demands rigorous and reliable evaluation. However, existing medical benchmarks remain static, suffering from two critical limitations: (1) data contamination, where test sets inadvertently leak into training corpora, leading to inflated performance estimates; and (2) temporal misalignment, failing to capture the rapid evolution of medical knowledge. Furthermore, current evaluation metrics for open-ended clinical reasoning often rely on either shallow lexical overlap (e.g., ROUGE) or subjective LLM-as-a-Judge scoring, both inadequate for verifying clinical correctness. To bridge these gaps, we introduce LiveMedBench, a continuously updated, contamination-free, and rubric-based benchmark that weekly harvests real-world clinical cases from online medical communities, ensuring strict temporal separation from model training data. We propose a Multi-Agent Clinical Curation Framework that filters raw data noise and validates clinical integrity against evidence-based medical principles. For evaluation, we develop an Automated Rubric-based Evaluation Framework that decomposes physician responses into granular, case-specific criteria, achieving substantially stronger alignment with expert physicians than LLM-as-a-Judge. To date, LiveMedBench comprises 2,756 real-world cases spanning 38 medical specialties and multiple languages, paired with 16,702 unique evaluation criteria. Extensive evaluation of 38 LLMs reveals that even the best-performing model achieves only 39.2%, and 84% of models exhibit performance degradation on post-cutoff cases, confirming pervasive data contamination risks. Error analysis further identifies contextual application-not factual knowledge-as the dominant bottleneck, with 35-48% of failures stemming from the inability to tailor medical knowledge to patient-specific constraints.</p> - https://arxiv.org/abs/2602.10367 - Tue, 10 Feb 2026 23:38:25 +0000 + Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments + https://arxiv.org/abs/2602.11964 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11964.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Romain Froger, Pierre Andrews, Matteo Bettini, Amar Budhiraja, Ricardo Silveira Cabral, Virginie Do, Emilien Garreau, Jean-Baptiste Gaya, Hugo Laurençon, Maxime Lecanu, Kunal Malkan, Dheeraj Mekala, Pierre Ménard, Gerard Moreno-Torres Bertran, Ulyana Piterbarg, Mikhail Plekhanov, Mathieu Rita, Andrey Rusakov, Vladislav Vorotilov, Mengjue Wang, Ian Yu, Amine Benhalloum, Grégoire Mialon, Thomas Scialom</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> We introduce Gaia2, a benchmark for evaluating large language model agents in realistic, asynchronous environments. Unlike prior static or synchronous evaluations, Gaia2 introduces scenarios where environments evolve independently of agent actions, requiring agents to operate under temporal constraints, adapt to noisy and dynamic events, resolve ambiguity, and collaborate with other agents. Each scenario is paired with a write-action verifier, enabling fine-grained, action-level evaluation and making Gaia2 directly usable for reinforcement learning from verifiable rewards. Our evaluation of state-of-the-art proprietary and open-source models shows that no model dominates across capabilities: GPT-5 (high) reaches the strongest overall score of 42% pass@1 but fails on time-sensitive tasks, Claude-4 Sonnet trades accuracy and speed for cost, Kimi-K2 leads among open-source models with 21% pass@1. These results highlight fundamental trade-offs between reasoning, efficiency, robustness, and expose challenges in closing the "sim2real" gap. Gaia2 is built on a consumer environment with the open-source Agents Research Environments platform and designed to be easy to extend. By releasing Gaia2 alongside the foundational ARE framework, we aim to provide the community with a flexible infrastructure for developing, benchmarking, and training the next generation of practical agent systems.</p> + https://arxiv.org/abs/2602.11964 + Thu, 12 Feb 2026 13:58:27 +0000 - When to Memorize and When to Stop: Gated Recurrent Memory for Long-Context Reasoning - https://arxiv.org/abs/2602.10560 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10560.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Leheng Sheng, Yongtao Zhang, Wenchang Ma, Yaorui Shi, Ting Huang, Xiang Wang, An Zhang, Ke Shen, Tat-Seng Chua</p><p><b>Upvotes:</b> 23</p><p><b>Summary:</b> While reasoning over long context is crucial for various real-world applications, it remains challenging for large language models (LLMs) as they suffer from performance degradation as the context length grows. Recent work MemAgent has tried to tackle this by processing context chunk-by-chunk in an RNN-like loop and updating a textual memory for final answering. However, this naive recurrent memory update faces two crucial drawbacks: (i) memory can quickly explode because it can update indiscriminately, even on evidence-free chunks; and (ii) the loop lacks an exit mechanism, leading to unnecessary computation after even sufficient evidence is collected. To address these issues, we propose GRU-Mem, which incorporates two text-controlled gates for more stable and efficient long-context reasoning. Specifically, in GRU-Mem, the memory only updates when the update gate is open and the recurrent loop will exit immediately once the exit gate is open. To endow the model with such capabilities, we introduce two reward signals r^{update} and r^{exit} within end-to-end RL, rewarding the correct updating and exiting behaviors respectively. Experiments on various long-context reasoning tasks demonstrate the effectiveness and efficiency of GRU-Mem, which generally outperforms the vanilla MemAgent with up to 400\% times inference speed acceleration.</p> - https://arxiv.org/abs/2602.10560 - Wed, 11 Feb 2026 06:14:53 +0000 + Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models + https://arxiv.org/abs/2602.12036 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12036.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xin Xu, Clive Bai, Kai Yang, Tianhao Chen, Yangkun Chen, Weijie Liu, Hao Chen, Yang Wang, Saiyong Yang, Can Yang</p><p><b>Upvotes:</b> 81</p><p><b>Summary:</b> Large-scale verifiable prompts underpin the success of Reinforcement Learning with Verifiable Rewards (RLVR), but they contain many uninformative examples and are costly to expand further. Recent studies focus on better exploiting limited training data by prioritizing hard prompts whose rollout pass rate is 0. However, easy prompts with a pass rate of 1 also become increasingly prevalent as training progresses, thereby reducing the effective data size. To mitigate this, we propose Composition-RL, a simple yet useful approach for better utilizing limited verifiable prompts targeting pass-rate-1 prompts. More specifically, Composition-RL automatically composes multiple problems into a new verifiable question and uses these compositional prompts for RL training. Extensive experiments across model sizes from 4B to 30B show that Composition-RL consistently improves reasoning capability over RL trained on the original dataset. Performance can be further boosted with a curriculum variant of Composition-RL that gradually increases compositional depth over training. Additionally, Composition-RL enables more effective cross-domain RL by composing prompts drawn from different domains. Codes, datasets, and models are available at https://github.com/XinXU-USTC/Composition-RL.</p> + https://arxiv.org/abs/2602.12036 + Thu, 12 Feb 2026 15:03:37 +0000 - Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters - https://arxiv.org/abs/2602.10604 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10604.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ailin Huang, Ang Li, Aobo Kong, Bin Wang, Binxing Jiao, Bo Dong, Bojun Wang, Boyu Chen, Brian Li, Buyun Ma, Chang Su, Changxin Miao, Changyi Wan, Chao Lou, Chen Hu, Chen Xu, Chenfeng Yu, Chengting Feng, Chengyuan Yao, Chunrui Han, Dan Ma, Dapeng Shi, Daxin Jiang, Dehua Ma, Deshan Sun, Di Qi, Enle Liu, Fajie Zhang, Fanqi Wan, Guanzhe Huang, Gulin Yan, Guoliang Cao, Guopeng Li, Han Cheng, Hangyu Guo, Hanshan Zhang, Hao Nie, Haonan Jia, Haoran Lv, Hebin Zhou, Hekun Lv, Heng Wang, Heung-Yeung Shum, Hongbo Huang, Hongbo Peng, Hongyu Zhou, Hongyuan Wang, Houyong Chen, Huangxi Zhu, Huimin Wu, Huiyong Guo, Jia Wang, Jian Zhou, Jianjian Sun, Jiaoren Wu, Jiaran Zhang, Jiashu Lv, Jiashuo Liu, Jiayi Fu, Jiayu Liu, Jie Cheng, Jie Luo, Jie Yang, Jie Zhou, Jieyi Hou, Jing Bai, Jingcheng Hu, Jingjing Xie, Jingwei Wu, Jingyang Zhang, Jishi Zhou, Junfeng Liu, Junzhe Lin, Ka Man Lo, Kai Liang, Kaibo Liu, Kaijun Tan, Kaiwen Yan, Kaixiang Li, Kang An, Kangheng Lin, Lei Yang, Liang Lv, Liang Zhao, Liangyu Chen, Lieyu Shi, Liguo Tan, Lin Lin, Lina Chen, Luck Ma, Mengqiang Ren, Michael Li, Ming Li, Mingliang Li, Mingming Zhang, Mingrui Chen, Mitt Huang, Na Wang, Peng Liu, Qi Han, Qian Zhao, Qinglin He, Qinxin Du, Qiuping Wu, Quan Sun, Rongqiu Yang, Ruihang Miao, Ruixin Han, Ruosi Wan, Ruyan Guo, Shan Wang, Shaoliang Pang, Shaowen Yang, Shengjie Fan, Shijie Shang, Shiliang Yang, Shiwei Li, Shuangshuang Tian, Siqi Liu, Siye Wu, Siyu Chen, Song Yuan, Tiancheng Cao, Tianchi Yue, Tianhao Cheng, Tianning Li, Tingdan Luo, Wang You, Wei Ji, Wei Yuan, Wei Zhang, Weibo Wu, Weihao Xie, Wen Sun, Wenjin Deng, Wenzhen Zheng, Wuxun Xie, Xiangfeng Wang, Xiangwen Kong, Xiangyu Liu, Xiangyu Zhang, Xiaobo Yang, Xiaojia Liu, Xiaolan Yuan, Xiaoran Jiao, Xiaoxiao Ren, Xiaoyun Zhang, Xin Li, Xin Liu, Xin Wu, Xing Chen, Xingping Yang, Xinran Wang, Xu Zhao, Xuan He, Xuanti Feng, Xuedan Cai, Xuqiang Zhou, Yanbo Yu, Yang Li, Yang Xu, Yanlin Lai, Yanming Xu, Yaoyu Wang, Yeqing Shen, Yibo Zhu, Yichen Lv, Yicheng Cao, Yifeng Gong, Yijing Yang, Yikun Yang, Yin Zhao, Yingxiu Zhao, Yinmin Zhang, Yitong Zhang, Yixuan Zhang, Yiyang Chen, Yongchi Zhao, Yongshen Long, Yongyao Wang, Yousong Guan, Yu Zhou, Yuang Peng, Yuanhao Ding, Yuantao Fan, Yuanzhen Yang, Yuchu Luo, Yudi Zhao, Yue Peng, Yueqiang Lin, Yufan Lu, Yuling Zhao, Yunzhou Ju, Yurong Zhang, Yusheng Li, Yuxiang Yang, Yuyang Chen, Yuzhu Cai, Zejia Weng, Zetao Hong, Zexi Li, Zhe Xie, Zheng Ge, Zheng Gong, Zheng Zeng, Zhenyi Lu, Zhewei Huang, Zhichao Chang, Zhiguo Huang, Zhiheng Hu, Zidong Yang, Zili Wang, Ziqi Ren, Zixin Zhang, Zixuan Wang</p><p><b>Upvotes:</b> 150</p><p><b>Summary:</b> We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.</p> - https://arxiv.org/abs/2602.10604 - Wed, 11 Feb 2026 07:53:51 +0000 + LawThinker: A Deep Research Legal Agent in Dynamic Environments + https://arxiv.org/abs/2602.12056 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12056.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xinyu Yang, Chenlong Deng, Tongyu Wen, Binyu Xie, Zhicheng Dou</p><p><b>Upvotes:</b> 31</p><p><b>Summary:</b> Legal reasoning requires not only correct outcomes but also procedurally compliant reasoning processes. However, existing methods lack mechanisms to verify intermediate reasoning steps, allowing errors such as inapplicable statute citations to propagate undetected through the reasoning chain. To address this, we propose LawThinker, an autonomous legal research agent that adopts an Explore-Verify-Memorize strategy for dynamic judicial environments. The core idea is to enforce verification as an atomic operation after every knowledge exploration step. A DeepVerifier module examines each retrieval result along three dimensions of knowledge accuracy, fact-law relevance, and procedural compliance, with a memory module for cross-round knowledge reuse in long-horizon tasks. Experiments on the dynamic benchmark J1-EVAL show that LawThinker achieves a 24% improvement over direct reasoning and an 11% gain over workflow-based methods, with particularly strong improvements on process-oriented metrics. Evaluations on three static benchmarks further confirm its generalization capability. The code is available at https://github.com/yxy-919/LawThinker-agent .</p> + https://arxiv.org/abs/2602.12056 + Thu, 12 Feb 2026 15:19:11 +0000 - Online Causal Kalman Filtering for Stable and Effective Policy Optimization - https://arxiv.org/abs/2602.10609 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10609.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shuo He, Lang Feng, Xin Cheng, Lei Feng, Bo An</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Reinforcement learning for large language models suffers from high-variance token-level importance sampling (IS) ratios, which would destabilize policy optimization at scale. To improve stability, recent methods typically use a fixed sequence-level IS ratio for all tokens in a sequence or adjust each token's IS ratio separately, thereby neglecting temporal off-policy derivation across tokens in a sequence. In this paper, we first empirically identify that local off-policy deviation is structurally inconsistent at the token level, which may distort policy-gradient updates across adjacent tokens and lead to training collapse. To address the issue, we propose Online Causal Kalman Filtering for stable and effective Policy Optimization (KPO). Concretely, we model the desired IS ratio as a latent state that evolves across tokens and apply a Kalman filter to update this state online and autoregressively based on the states of past tokens, regardless of future tokens. The resulting filtered IS ratios preserve token-wise local structure-aware variation while strongly smoothing noise spikes, yielding more stable and effective policy updates. Experimentally, KPO achieves superior results on challenging math reasoning datasets compared with state-of-the-art counterparts.</p> - https://arxiv.org/abs/2602.10609 - Wed, 11 Feb 2026 07:57:43 +0000 + DeepSight: An All-in-One LM Safety Toolkit + https://arxiv.org/abs/2602.12092 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12092.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Bo Zhang, Jiaxuan Guo, Lijun Li, Dongrui Liu, Sujin Chen, Guanxu Chen, Zhijie Zheng, Qihao Lin, Lewen Yan, Chen Qian, Yijin Zhou, Yuyao Wu, Shaoxiong Guo, Tianyi Du, Jingyi Yang, Xuhao Hu, Ziqi Miao, Xiaoya Lu, Jing Shao, Xia Hu</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> As the development of Large Models (LMs) progresses rapidly, their safety is also a priority. In current Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) safety workflow, evaluation, diagnosis, and alignment are often handled by separate tools. Specifically, safety evaluation can only locate external behavioral risks but cannot figure out internal root causes. Meanwhile, safety diagnosis often drifts from concrete risk scenarios and remains at the explainable level. In this way, safety alignment lack dedicated explanations of changes in internal mechanisms, potentially degrading general capabilities. To systematically address these issues, we propose an open-source project, namely DeepSight, to practice a new safety evaluation-diagnosis integrated paradigm. DeepSight is low-cost, reproducible, efficient, and highly scalable large-scale model safety evaluation project consisting of a evaluation toolkit DeepSafe and a diagnosis toolkit DeepScan. By unifying task and data protocols, we build a connection between the two stages and transform safety evaluation from black-box to white-box insight. Besides, DeepSight is the first open source toolkit that support the frontier AI risk evaluation and joint safety evaluation and diagnosis.</p> + https://arxiv.org/abs/2602.12092 + Thu, 12 Feb 2026 15:43:14 +0000 - How Do Decoder-Only LLMs Perceive Users? Rethinking Attention Masking for User Representation Learning - https://arxiv.org/abs/2602.10622 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10622.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiahao Yuan, Yike Xu, Jinyong Wen, Baokun Wang, Yang Chen, Xiaotong Lin, Wuliang Huang, Ziyi Gao, Xing Fu, Yu Cheng, Weiqiang Wang</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> Decoder-only large language models are increasingly used as behavioral encoders for user representation learning, yet the impact of attention masking on the quality of user embeddings remains underexplored. In this work, we conduct a systematic study of causal, hybrid, and bidirectional attention masks within a unified contrastive learning framework trained on large-scale real-world Alipay data that integrates long-horizon heterogeneous user behaviors. To improve training dynamics when transitioning from causal to bidirectional attention, we propose Gradient-Guided Soft Masking, a gradient-based pre-warmup applied before a linear scheduler that gradually opens future attention during optimization. Evaluated on 9 industrial user cognition benchmarks covering prediction, preference, and marketing sensitivity tasks, our approach consistently yields more stable training and higher-quality bidirectional representations compared with causal, hybrid, and scheduler-only baselines, while remaining compatible with decoder pretraining. Overall, our findings highlight the importance of masking design and training transition in adapting decoder-only LLMs for effective user representation learning. Our code is available at https://github.com/JhCircle/Deepfind-GGSM.</p> - https://arxiv.org/abs/2602.10622 - Wed, 11 Feb 2026 08:12:43 +0000 + GigaBrain-0.5M*: a VLA That Learns From World Model-Based Reinforcement Learning + https://arxiv.org/abs/2602.12099 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12099.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> GigaBrain Team, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Jie Li, Jindi Lv, Jingyu Liu, Lv Feng, Mingming Yu, Peng Li, Qiuping Deng, Tianze Liu, Xinyu Zhou, Xinze Chen, Xiaofeng Wang, Yang Wang, Yifan Li, Yifei Nie, Yilong Li, Yukun Zhou, Yun Ye, Zhichao Liu, Zheng Zhu</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> Vision-language-action (VLA) models that directly predict multi-step action chunks from current observations face inherent limitations due to constrained scene understanding and weak future anticipation capabilities. In contrast, video world models pre-trained on web-scale video corpora exhibit robust spatiotemporal reasoning and accurate future prediction, making them a natural foundation for enhancing VLA learning. Therefore, we propose GigaBrain-0.5M*, a VLA model trained via world model-based reinforcement learning. Built upon GigaBrain-0.5, which is pre-trained on over 10,000 hours of robotic manipulation data, whose intermediate version currently ranks first on the international RoboChallenge benchmark. GigaBrain-0.5M* further integrates world model-based reinforcement learning via RAMP (Reinforcement leArning via world Model-conditioned Policy) to enable robust cross-task adaptation. Empirical results demonstrate that RAMP achieves substantial performance gains over the RECAP baseline, yielding improvements of approximately 30\% on challenging tasks including Laundry Folding, Box Packing, and Espresso Preparation. Critically, GigaBrain-0.5M^* exhibits reliable long-horizon execution, consistently accomplishing complex manipulation tasks without failure as validated by real-world deployment videos on our https://gigabrain05m.github.io{project page}.</p> + https://arxiv.org/abs/2602.12099 + Thu, 12 Feb 2026 15:55:19 +0000 - UMEM: Unified Memory Extraction and Management Framework for Generalizable Memory - https://arxiv.org/abs/2602.10652 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10652.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yongshi Ye, Hui Jiang, Feihu Jiang, Tian Lan, Yichao Du, Biao Fu, Xiaodong Shi, Qianghuai Jia, Longyue Wang, Weihua Luo</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Self-evolving memory serves as the trainable parameters for Large Language Models (LLMs)-based agents, where extraction (distilling insights from experience) and management (updating the memory bank) must be tightly coordinated. Existing methods predominately optimize memory management while treating memory extraction as a static process, resulting in poor generalization, where agents accumulate instance-specific noise rather than robust memories. To address this, we propose Unified Memory Extraction and Management (UMEM), a self-evolving agent framework that jointly optimizes a Large Language Model to simultaneous extract and manage memories. To mitigate overfitting to specific instances, we introduce Semantic Neighborhood Modeling and optimize the model with a neighborhood-level marginal utility reward via GRPO. This approach ensures memory generalizability by evaluating memory utility across clusters of semantically related queries. Extensive experiments across five benchmarks demonstrate that UMEM significantly outperforms highly competitive baselines, achieving up to a 10.67% improvement in multi-turn interactive tasks. Futhermore, UMEM maintains a monotonic growth curve during continuous evolution. Codes and models will be publicly released.</p> - https://arxiv.org/abs/2602.10652 - Wed, 11 Feb 2026 08:58:41 +0000 + P-GenRM: Personalized Generative Reward Model with Test-time User-based Scaling + https://arxiv.org/abs/2602.12116 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12116.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Pinyi Zhang, Ting-En Lin, Yuchuan Wu, Jingyang Chen, Zongqi Wang, Hua Yang, Ze Xu, Fei Huang, Kai Zhang, Yongbin Li</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Personalized alignment of large language models seeks to adapt responses to individual user preferences, typically via reinforcement learning. A key challenge is obtaining accurate, user-specific reward signals in open-ended scenarios. Existing personalized reward models face two persistent limitations: (1) oversimplifying diverse, scenario-specific preferences into a small, fixed set of evaluation principles, and (2) struggling with generalization to new users with limited feedback. To this end, we propose P-GenRM, the first Personalized Generative Reward Model with test-time user-based scaling. P-GenRM transforms preference signals into structured evaluation chains that derive adaptive personas and scoring rubrics across various scenarios. It further clusters users into User Prototypes and introduces a dual-granularity scaling mechanism: at the individual level, it adaptively scales and aggregates each user's scoring scheme; at the prototype level, it incorporates preferences from similar users. This design mitigates noise in inferred preferences and enhances generalization to unseen users through prototype-based transfer. Empirical results show that P-GenRM achieves state-of-the-art results on widely-used personalized reward model benchmarks, with an average improvement of 2.31%, and demonstrates strong generalization on an out-of-distribution dataset. Notably, Test-time User-based scaling provides an additional 3% boost, demonstrating stronger personalized alignment with test-time scalability.</p> + https://arxiv.org/abs/2602.12116 + Thu, 12 Feb 2026 16:07:22 +0000 - Spend Search Where It Pays: Value-Guided Structured Sampling and Optimization for Generative Recommendation - https://arxiv.org/abs/2602.10699 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10699.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jie Jiang, Yangru Huang, Zeyu Wang, Changping Wang, Yuling Xiong, Jun Zhang, Huan Yu</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Generative recommendation via autoregressive models has unified retrieval and ranking into a single conditional generation framework. However, fine-tuning these models with Reinforcement Learning (RL) often suffers from a fundamental probability-reward mismatch. Conventional likelihood-dominated decoding (e.g., beam search) exhibits a myopic bias toward locally probable prefixes, which causes two critical failures: (1) insufficient exploration, where high-reward items in low-probability branches are prematurely pruned and rarely sampled, and (2) advantage compression, where trajectories sharing high-probability prefixes receive highly correlated rewards with low within-group variance, yielding a weak comparative signal for RL. To address these challenges, we propose V-STAR, a Value-guided Sampling and Tree-structured Advantage Reinforcement framework. V-STAR forms a self-evolving loop via two synergistic components. First, a Value-Guided Efficient Decoding (VED) is developed to identify decisive nodes and selectively deepen high-potential prefixes. This improves exploration efficiency without exhaustive tree search. Second, we propose Sibling-GRPO, which exploits the induced tree topology to compute sibling-relative advantages and concentrates learning signals on decisive branching decisions. Extensive experiments on both offline and online datasets demonstrate that V-STAR outperforms state-of-the-art baselines, delivering superior accuracy and candidate-set diversity under strict latency constraints.</p> - https://arxiv.org/abs/2602.10699 - Wed, 11 Feb 2026 09:57:36 +0000 + Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation + https://arxiv.org/abs/2602.12125 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12125.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wenkai Yang, Weijie Liu, Ruobing Xie, Kai Yang, Saiyong Yang, Yankai Lin</p><p><b>Upvotes:</b> 53</p><p><b>Summary:</b> On-policy distillation (OPD), which aligns the student with the teacher's logit distribution on student-generated trajectories, has demonstrated strong empirical gains in improving student performance and often outperforms off-policy distillation and reinforcement learning (RL) paradigms. In this work, we first theoretically show that OPD is a special case of dense KL-constrained RL where the reward function and the KL regularization are always weighted equally and the reference model can by any model. Then, we propose the Generalized On-Policy Distillation (G-OPD) framework, which extends the standard OPD objective by introducing a flexible reference model and a reward scaling factor that controls the relative weight of the reward term against the KL regularization. Through comprehensive experiments on math reasoning and code generation tasks, we derive two novel insights: (1) Setting the reward scaling factor to be greater than 1 (i.e., reward extrapolation), which we term ExOPD, consistently improves over standard OPD across a range of teacher-student size pairings. In particular, in the setting where we merge the knowledge from different domain experts, obtained by applying domain-specific RL to the same student model, back into the original student, ExOPD enables the student to even surpass the teacher's performance boundary and outperform the domain teachers. (2) Building on ExOPD, we further find that in the strong-to-weak distillation setting (i.e., distilling a smaller student from a larger teacher), performing reward correction by choosing the reference model as the teacher's base model before RL yields a more accurate reward signal and further improves distillation performance. However, this choice assumes access to the teacher's pre-RL variant and incurs more computational overhead. We hope our work offers new insights for future research on OPD.</p> + https://arxiv.org/abs/2602.12125 + Thu, 12 Feb 2026 16:14:29 +0000 - Benchmarking Large Language Models for Knowledge Graph Validation - https://arxiv.org/abs/2602.10748 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10748.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Farzad Shami, Stefano Marchesin, Gianmaria Silvello</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Knowledge Graphs (KGs) store structured factual knowledge by linking entities through relationships, crucial for many applications. These applications depend on the KG's factual accuracy, so verifying facts is essential, yet challenging. Expert manual verification is ideal but impractical on a large scale. Automated methods show promise but are not ready for real-world KGs. Large Language Models (LLMs) offer potential with their semantic understanding and knowledge access, yet their suitability and effectiveness for KG fact validation remain largely unexplored. In this paper, we introduce FactCheck, a benchmark designed to evaluate LLMs for KG fact validation across three key dimensions: (1) LLMs internal knowledge; (2) external evidence via Retrieval-Augmented Generation (RAG); and (3) aggregated knowledge employing a multi-model consensus strategy. We evaluated open-source and commercial LLMs on three diverse real-world KGs. FactCheck also includes a RAG dataset with 2+ million documents tailored for KG fact validation. Additionally, we offer an interactive exploration platform for analyzing verification decisions. The experimental analyses demonstrate that while LLMs yield promising results, they are still not sufficiently stable and reliable to be used in real-world KG validation scenarios. Integrating external evidence through RAG methods yields fluctuating performance, providing inconsistent improvements over more streamlined approaches -- at higher computational costs. Similarly, strategies based on multi-model consensus do not consistently outperform individual models, underscoring the lack of a one-fits-all solution. These findings further emphasize the need for a benchmark like FactCheck to systematically evaluate and drive progress on this difficult yet crucial task.</p> - https://arxiv.org/abs/2602.10748 - Wed, 11 Feb 2026 11:24:46 +0000 + dVoting: Fast Voting for dLLMs + https://arxiv.org/abs/2602.12153 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12153.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sicheng Feng, Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang</p><p><b>Upvotes:</b> 17</p><p><b>Summary:</b> Diffusion Large Language Models (dLLMs) represent a new paradigm beyond autoregressive modeling, offering competitive performance while naturally enabling a flexible decoding process. Specifically, dLLMs can generate tokens at arbitrary positions in parallel, endowing them with significant potential for parallel test-time scaling, which was previously constrained by severe inefficiency in autoregressive modeling. In this work, we introduce dVoting, a fast voting technique that boosts reasoning capability without training, with only an acceptable extra computational overhead. dVoting is motivated by the observation that, across multiple samples for the same prompt, token predictions remain largely consistent, whereas performance is determined by a small subset of tokens exhibiting cross-sample variability. Leveraging the arbitrary-position generation capability of dLLMs, dVoting performs iterative refinement by sampling, identifying uncertain tokens via consistency analysis, regenerating them through voting, and repeating this process until convergence. Extensive evaluations demonstrate that dVoting consistently improves performance across various benchmarks. It achieves gains of 6.22%-7.66% on GSM8K, 4.40%-7.20% on MATH500, 3.16%-14.84% on ARC-C, and 4.83%-5.74% on MMLU. Our code is available at https://github.com/fscdc/dVoting</p> + https://arxiv.org/abs/2602.12153 + Thu, 12 Feb 2026 16:35:05 +0000 - GoodVibe: Security-by-Vibe for LLM-Based Code Generation - https://arxiv.org/abs/2602.10778 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10778.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Maximilian Thang, Lichao Wu, Sasha Behrouzi, Mohamadreza Rostami, Jona te Lintelo, Stjepan Picek, Ahmad-Reza Sadeghi</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Large language models (LLMs) are increasingly used for code generation in fast, informal development workflows, often referred to as vibe coding, where speed and convenience are prioritized, and security requirements are rarely made explicit. In this setting, models frequently produce functionally correct but insecure code, creating a growing security risk. Existing approaches to improving code security rely on full-parameter fine-tuning or parameter-efficient adaptations, which are either costly and prone to catastrophic forgetting or operate at coarse granularity with limited interpretability and control. We present GoodVibe, a neuron-level framework for improving the security of code language models by default. GoodVibe is based on the key insight that security-relevant reasoning is localized to a small subset of neurons. We identify these neurons using gradient-based attribution from a supervised security task and perform neuron-selective fine-tuning that updates only this security-critical subspace. To further reduce training cost, we introduce activation-driven neuron clustering, enabling structured updates with minimal overhead. We evaluate GoodVibe on six LLMs across security-critical programming languages, including C++, Java, Swift, and Go. GoodVibe substantially improves the security of generated code while preserving general model utility, achieving up to a 2.5x improvement over base models, matching or exceeding full fine-tuning with over 4,700x fewer trainable parameters, and reducing training computation by more than 3.6x compared to the parameter-efficient baseline (LoRA). Our results demonstrate that neuron-level optimization offers an effective and scalable approach to securing code generation without sacrificing efficiency or generality.</p> - https://arxiv.org/abs/2602.10778 - Wed, 11 Feb 2026 12:10:14 +0000 + Sci-CoE: Co-evolving Scientific Reasoning LLMs via Geometric Consensus with Sparse Supervision + https://arxiv.org/abs/2602.12164 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12164.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiaohan He, Shiyang Feng, Songtao Huang, Lei Bai, Bin Wang, Bo Zhang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Large language models (LLMs) have demonstrated exceptional reasoning capabilities, and co-evolving paradigms have shown promising results in domains such as code and math. However, in scientific reasoning tasks, these models remain fragile due to unreliable solution evaluation and limited diversity in verification strategies. In this work, we propose Sci-CoE, a two-stage scientific co-evolving framework that enables models to self-evolve as both solver and verifier through a transition from sparse supervision to unsupervised learning. In the first stage, the model uses a small set of annotated data to establish fundamental correctness judgment anchors for the Verifier. In the second stage, we introduce a geometric reward mechanism that jointly considers consensus, reliability, and diversity, driving large-scale self-iteration on unlabeled data. Experiments on several general scientific benchmarks demonstrate that Sci-CoE enhances complex reasoning capabilities and exhibits strong scalability, facilitating the construction of more robust and diverse evaluation systems. Codes are available at https://github.com/InternScience/Sci-CoE.</p> + https://arxiv.org/abs/2602.12164 + Thu, 12 Feb 2026 16:46:00 +0000 - FedPS: Federated data Preprocessing via aggregated Statistics - https://arxiv.org/abs/2602.10870 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10870.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xuefeng Xu, Graham Cormode</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Federated Learning (FL) enables multiple parties to collaboratively train machine learning models without sharing raw data. However, before training, data must be preprocessed to address missing values, inconsistent formats, and heterogeneous feature scales. This preprocessing stage is critical for model performance but is largely overlooked in FL research. In practical FL systems, privacy constraints prohibit centralizing raw data, while communication efficiency introduces further challenges for distributed preprocessing. We introduce FedPS, a unified framework for federated data preprocessing based on aggregated statistics. FedPS leverages data-sketching techniques to efficiently summarize local datasets while preserving essential statistical information. Building on these summaries, we design federated algorithms for feature scaling, encoding, discretization, and missing-value imputation, and extend preprocessing-related models such as k-Means, k-Nearest Neighbors, and Bayesian Linear Regression to both horizontal and vertical FL settings. FedPS provides flexible, communication-efficient, and consistent preprocessing pipelines for practical FL deployments.</p> - https://arxiv.org/abs/2602.10870 - Wed, 11 Feb 2026 13:58:55 +0000 + Single-minus gluon tree amplitudes are nonzero + https://arxiv.org/abs/2602.12176 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12176.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Alfredo Guevara, Alexandru Lupsasca, David Skinner, Andrew Strominger, Kevin Weil</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Single-minus tree-level n-gluon scattering amplitudes are reconsidered. Often presumed to vanish, they are shown here to be nonvanishing for certain "half-collinear" configurations existing in Klein space or for complexified momenta. We derive a piecewise-constant closed-form expression for the decay of a single minus-helicity gluon into n-1 plus-helicity gluons as a function of their momenta. This formula nontrivially satisfies multiple consistency conditions including Weinberg's soft theorem.</p> + https://arxiv.org/abs/2602.12176 + Thu, 12 Feb 2026 17:09:06 +0000 - FeatureBench: Benchmarking Agentic Coding for Complex Feature Development - https://arxiv.org/abs/2602.10975 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10975.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qixing Zhou, Jiacheng Zhang, Haiyang Wang, Rui Hao, Jiahe Wang, Minghao Han, Yuxue Yang, Shuzhe Wu, Feiyang Pan, Lue Fan, Dandan Tu, Zhaoxiang Zhang</p><p><b>Upvotes:</b> 17</p><p><b>Summary:</b> Agents powered by large language models (LLMs) are increasingly adopted in the software industry, contributing code as collaborators or even autonomous developers. As their presence grows, it becomes important to assess the current boundaries of their coding abilities. Existing agentic coding benchmarks, however, cover a limited task scope, e.g., bug fixing within a single pull request (PR), and often rely on non-executable evaluations or lack an automated approach for continually updating the evaluation coverage. To address such issues, we propose FeatureBench, a benchmark designed to evaluate agentic coding performance in end-to-end, feature-oriented software development. FeatureBench incorporates an execution-based evaluation protocol and a scalable test-driven method that automatically derives tasks from code repositories with minimal human effort. By tracing from unit tests along a dependency graph, our approach can identify feature-level coding tasks spanning multiple commits and PRs scattered across the development timeline, while ensuring the proper functioning of other features after the separation. Using this framework, we curated 200 challenging evaluation tasks and 3825 executable environments from 24 open-source repositories in the first version of our benchmark. Empirical evaluation reveals that the state-of-the-art agentic model, such as Claude 4.5 Opus, which achieves a 74.4% resolved rate on SWE-bench, succeeds on only 11.0% of tasks, opening new opportunities for advancing agentic coding. Moreover, benefiting from our automated task collection toolkit, FeatureBench can be easily scaled and updated over time to mitigate data leakage. The inherent verifiability of constructed environments also makes our method potentially valuable for agent training.</p> - https://arxiv.org/abs/2602.10975 - Wed, 11 Feb 2026 16:06:32 +0000 + ExStrucTiny: A Benchmark for Schema-Variable Structured Information Extraction from Document Images + https://arxiv.org/abs/2602.12203 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12203.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Mathieu Sibue, Andres Muñoz Garza, Samuel Mensah, Pranav Shetty, Zhiqiang Ma, Xiaomo Liu, Manuela Veloso</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Enterprise documents, such as forms and reports, embed critical information for downstream applications like data archiving, automated workflows, and analytics. Although generalist Vision Language Models (VLMs) perform well on established document understanding benchmarks, their ability to conduct holistic, fine-grained structured extraction across diverse document types and flexible schemas is not well studied. Existing Key Entity Extraction (KEE), Relation Extraction (RE), and Visual Question Answering (VQA) datasets are limited by narrow entity ontologies, simple queries, or homogeneous document types, often overlooking the need for adaptable and structured extraction. To address these gaps, we introduce ExStrucTiny, a new benchmark dataset for structured Information Extraction (IE) from document images, unifying aspects of KEE, RE, and VQA. Built through a novel pipeline combining manual and synthetic human-validated samples, ExStrucTiny covers more varied document types and extraction scenarios. We analyze open and closed VLMs on this benchmark, highlighting challenges such as schema adaptation, query under-specification, and answer localization. We hope our work provides a bedrock for improving generalist models for structured IE in documents.</p> + https://arxiv.org/abs/2602.12203 + Thu, 12 Feb 2026 17:38:57 +0000 - CLI-Gym: Scalable CLI Task Generation via Agentic Environment Inversion - https://arxiv.org/abs/2602.10999 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.10999.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yusong Lin, Haiyang Wang, Shuzhe Wu, Lue Fan, Feiyang Pan, Sanyuan Zhao, Dandan Tu</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Agentic coding requires agents to effectively interact with runtime environments, e.g., command line interfaces (CLI), so as to complete tasks like resolving dependency issues, fixing system problems, etc. But it remains underexplored how such environment-intensive tasks can be obtained at scale to enhance agents' capabilities. To address this, based on an analogy between the Dockerfile and the agentic task, we propose to employ agents to simulate and explore environment histories, guided by execution feedback. By tracing histories of a healthy environment, its state can be inverted to an earlier one with runtime failures, from which a task can be derived by packing the buggy state and the corresponding error messages. With our method, named CLI-Gym, a total of 1,655 environment-intensive tasks are derived, being the largest collection of its kind. Moreover, with curated successful trajectories, our fine-tuned model, named LiberCoder, achieves substantial absolute improvements of +21.1% (to 46.1%) on Terminal-Bench, outperforming various strong baselines. To our knowledge, this is the first public pipeline for scalable derivation of environment-intensive tasks.</p> - https://arxiv.org/abs/2602.10999 - Wed, 11 Feb 2026 16:22:18 +0000 + DeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing + https://arxiv.org/abs/2602.12205 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12205.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dianyi Wang, Ruihang Li, Feng Han, Chaofan Ma, Wei Song, Siyuan Wang, Yibin Wang, Yi Xin, Hongjian Liu, Zhixiong Zhang, Shengyuan Ding, Tianhang Wang, Zhenglin Cheng, Tao Lin, Cheng Jin, Kaicheng Yu, Jingjing Chen, Wenjie Wang, Zhongyu Wei, Jiaqi Wang</p><p><b>Upvotes:</b> 60</p><p><b>Summary:</b> Current unified multimodal models for image generation and editing typically rely on massive parameter scales (e.g., >10B), entailing prohibitive training costs and deployment footprints. In this work, we present DeepGen 1.0, a lightweight 5B unified model that achieves comprehensive capabilities competitive with or surpassing much larger counterparts. To overcome the limitations of compact models in semantic understanding and fine-grained control, we introduce Stacked Channel Bridging (SCB), a deep alignment framework that extracts hierarchical features from multiple VLM layers and fuses them with learnable 'think tokens' to provide the generative backbone with structured, reasoning-rich guidance. We further design a data-centric training strategy spanning three progressive stages: (1) Alignment Pre-training on large-scale image-text pairs and editing triplets to synchronize VLM and DiT representations, (2) Joint Supervised Fine-tuning on a high-quality mixture of generation, editing, and reasoning tasks to foster omni-capabilities, and (3) Reinforcement Learning with MR-GRPO, which leverages a mixture of reward functions and supervision signals, resulting in substantial gains in generation quality and alignment with human preferences, while maintaining stable training progress and avoiding visual artifacts. Despite being trained on only ~50M samples, DeepGen 1.0 achieves leading performance across diverse benchmarks, surpassing the 80B HunyuanImage by 28% on WISE and the 27B Qwen-Image-Edit by 37% on UniREditBench. By open-sourcing our training code, weights, and datasets, we provide an efficient, high-performance alternative to democratize unified multimodal research.</p> + https://arxiv.org/abs/2602.12205 + Thu, 12 Feb 2026 17:44:24 +0000 - ROCKET: Rapid Optimization via Calibration-guided Knapsack Enhanced Truncation for Efficient Model Compression - https://arxiv.org/abs/2602.11008 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11008.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ammar Ali, Baher Mohammad, Denis Makhov, Dmitriy Shopkhoev, Magauiya Zhussip, Stamatios Lefkimmiatis</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> We present ROCKET, a training-free model compression method that achieves state-of-the-art performance in comparison with factorization, structured-sparsification and dynamic compression baselines. Operating under a global compression budget, ROCKET comprises two key innovations: First, it formulates layer-wise compression allocation as a multi-choice knapsack problem, selecting the optimal compression level for each layer to minimize total reconstruction error while adhering to a target model size. Second, it introduces a single-step sparse matrix factorization inspired by dictionary learning: using only a small calibration set, it sparsifies weight coefficients based on activation-weights sensitivity and then updates the dictionary in closed form via least squares bypassing iterative optimization, sparse coding, or backpropagation entirely. ROCKET consistently outperforms existing compression approaches across different model architectures at 20-50\% compression rates. Notably, it retains over 90\% of the original model's performance at 30\% compression without any fine-tuning. Moreover, when applying a light fine-tuning phase, recovery is substantially enhanced: for instance, compressing Qwen3-14B to an 8B-parameter model and healing it with just 30 million tokens yields performance nearly on par with the original Qwen3-8B. The code for ROCKET is at github.com/mts-ai/ROCKET/tree/main.</p> - https://arxiv.org/abs/2602.11008 - Wed, 11 Feb 2026 16:34:52 +0000 + T3D: Few-Step Diffusion Language Models via Trajectory Self-Distillation with Direct Discriminative Optimization + https://arxiv.org/abs/2602.12262 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12262.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tunyu Zhang, Xinxi Zhang, Ligong Han, Haizhou Shi, Xiaoxiao He, Zhuowei Li, Hao Wang, Kai Xu, Akash Srivastava, Hao Wang, Vladimir Pavlovic, Dimitris N. Metaxas</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Diffusion large language models (DLLMs) have the potential to enable fast text generation by decoding multiple tokens in parallel. However, in practice, their inference efficiency is constrained by the need for many refinement steps, while aggressively reducing the number of steps leads to a substantial degradation in generation quality. To alleviate this, we propose a trajectory self-distillation framework that improves few-step decoding by distilling the model's own generative trajectories. We incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that promotes mode-seeking distillation and encourages the student to concentrate on high-probability teacher modes. Across benchmarks, our approach consistently outperforms strong few-step baselines and standard training under tight step budgets. Although full-step decoding remains superior, we substantially narrow the gap, establishing a strong foundation towards practical few-step DLLMs. The source code is available at https://github.com/Tyrion58/T3D.</p> + https://arxiv.org/abs/2602.12262 + Thu, 12 Feb 2026 18:52:35 +0000 - DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning - https://arxiv.org/abs/2602.11089 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11089.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yicheng Chen, Zerun Ma, Xinchen Xie, Yining Li, Kai Chen</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> In the current landscape of Large Language Models (LLMs), the curation of large-scale, high-quality training data is a primary driver of model performance. A key lever is the data recipe, which comprises a data processing pipeline to transform raw sources into training corpora. Despite the growing use of LLMs to automate individual data processing steps, such as data synthesis and filtering, the overall design of data recipes remains largely manual and labor-intensive, requiring substantial human expertise and iteration. To bridge this gap, we formulate end-to-end data recipe generation for LLM adaptation. Given a target benchmark and a pool of available data sources, a model is required to output a complete data recipe that adapts a base LLM to the target task. We present DataChef-32B, which performs online reinforcement learning using a proxy reward that predicts downstream performance for candidate recipes. Across six held-out tasks, DataChef-32B produces practical recipes that reach comparable downstream performance to those curated by human experts. Notably, the recipe from DataChef-32B adapts Qwen3-1.7B-Base to the math domain, achieving 66.7 on AIME'25 and surpassing Qwen3-1.7B. This work sheds new light on automating LLM training and developing self-evolving AI systems.</p> - https://arxiv.org/abs/2602.11089 - Wed, 11 Feb 2026 17:56:15 +0000 - - - GameDevBench: Evaluating Agentic Capabilities Through Game Development - https://arxiv.org/abs/2602.11103 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11103.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wayne Chi, Yixiong Fang, Arnav Yayavaram, Siddharth Yayavaram, Seth Karten, Qiuhong Anna Wei, Runkun Chen, Alexander Wang, Valerie Chen, Ameet Talwalkar, Chris Donahue</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Despite rapid progress on coding agents, progress on their multimodal counterparts has lagged behind. A key challenge is the scarcity of evaluation testbeds that combine the complexity of software development with the need for deep multimodal understanding. Game development provides such a testbed as agents must navigate large, dense codebases while manipulating intrinsically multimodal assets such as shaders, sprites, and animations within a visual game scene. We present GameDevBench, the first benchmark for evaluating agents on game development tasks. GameDevBench consists of 132 tasks derived from web and video tutorials. Tasks require significant multimodal understanding and are complex -- the average solution requires over three times the amount of lines of code and file changes compared to prior software development benchmarks. Agents still struggle with game development, with the best agent solving only 54.5% of tasks. We find a strong correlation between perceived task difficulty and multimodal complexity, with success rates dropping from 46.9% on gameplay-oriented tasks to 31.6% on 2D graphics tasks. To improve multimodal capability, we introduce two simple image and video-based feedback mechanisms for agents. Despite their simplicity, these methods consistently improve performance, with the largest change being an increase in Claude Sonnet 4.5's performance from 33.3% to 47.7%. We release GameDevBench publicly to support further research into agentic game development.</p> - https://arxiv.org/abs/2602.11103 - Wed, 11 Feb 2026 18:15:11 +0000 - - - PhyCritic: Multimodal Critic Models for Physical AI - https://arxiv.org/abs/2602.11124 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11124.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tianyi Xiong, Shihao Wang, Guilin Liu, Yi Dong, Ming Li, Heng Huang, Jan Kautz, Zhiding Yu</p><p><b>Upvotes:</b> 43</p><p><b>Summary:</b> With the rapid development of large multimodal models, reliable judge and critic models have become essential for open-ended evaluation and preference alignment, providing pairwise preferences, numerical scores, and explanatory justifications for assessing model-generated responses. However, existing critics are primarily trained in general visual domains such as captioning or image question answering, leaving physical AI tasks involving perception, causal reasoning, and planning largely underexplored. We introduce PhyCritic, a multimodal critic model optimized for physical AI through a two-stage RLVR pipeline: a physical skill warmup stage that enhances physically oriented perception and reasoning, followed by self-referential critic finetuning, where the critic generates its own prediction as an internal reference before judging candidate responses, improving judgment stability and physical correctness. Across both physical and general-purpose multimodal judge benchmarks, PhyCritic achieves strong performance gains over open-source baselines and, when applied as a policy model, further improves perception and reasoning in physically grounded tasks.</p> - https://arxiv.org/abs/2602.11124 - Wed, 11 Feb 2026 18:35:39 +0000 - - - Weight Decay Improves Language Model Plasticity - https://arxiv.org/abs/2602.11137 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11137.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tessa Han, Sebastian Bordt, Hanlin Zhang, Sham Kakade</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> The prevailing paradigm in large language model (LLM) development is to pretrain a base model, then perform further training to improve performance and model behavior. However, hyperparameter optimization and scaling laws have been studied primarily from the perspective of the base model's validation loss, ignoring downstream adaptability. In this work, we study pretraining from the perspective of model plasticity, that is, the ability of the base model to successfully adapt to downstream tasks through fine-tuning. We focus on the role of weight decay, a key regularization parameter during pretraining. Through systematic experiments, we show that models trained with larger weight decay values are more plastic, meaning they show larger performance gains when fine-tuned on downstream tasks. This phenomenon can lead to counterintuitive trade-offs where base models that perform worse after pretraining can perform better after fine-tuning. Further investigation of weight decay's mechanistic effects on model behavior reveals that it encourages linearly separable representations, regularizes attention matrices, and reduces overfitting on the training data. In conclusion, this work demonstrates the importance of using evaluation metrics beyond cross-entropy loss for hyperparameter optimization and casts light on the multifaceted role of that a single optimization hyperparameter plays in shaping model behavior.</p> - https://arxiv.org/abs/2602.11137 - Wed, 11 Feb 2026 18:49:26 +0000 - - - GENIUS: Generative Fluid Intelligence Evaluation Suite - https://arxiv.org/abs/2602.11144 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11144.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ruichuan An, Sihan Yang, Ziyu Guo, Wei Dai, Zijun Shen, Haodong Li, Renrui Zhang, Xinyu Wei, Guopeng Li, Wenshan Wu, Wentao Zhang</p><p><b>Upvotes:</b> 42</p><p><b>Summary:</b> Unified Multimodal Models (UMMs) have shown remarkable progress in visual generation. Yet, existing benchmarks predominantly assess Crystallized Intelligence, which relies on recalling accumulated knowledge and learned schemas. This focus overlooks Generative Fluid Intelligence (GFI): the capacity to induce patterns, reason through constraints, and adapt to novel scenarios on the fly. To rigorously assess this capability, we introduce GENIUS (GEN Fluid Intelligence EvalUation Suite). We formalize GFI as a synthesis of three primitives. These include Inducing Implicit Patterns (e.g., inferring personalized visual preferences), Executing Ad-hoc Constraints (e.g., visualizing abstract metaphors), and Adapting to Contextual Knowledge (e.g., simulating counter-intuitive physics). Collectively, these primitives challenge models to solve problems grounded entirely in the immediate context. Our systematic evaluation of 12 representative models reveals significant performance deficits in these tasks. Crucially, our diagnostic analysis disentangles these failure modes. It demonstrates that deficits stem from limited context comprehension rather than insufficient intrinsic generative capability. To bridge this gap, we propose a training-free attention intervention strategy. Ultimately, GENIUS establishes a rigorous standard for GFI, guiding the field beyond knowledge utilization toward dynamic, general-purpose reasoning. Our dataset and code will be released at: https://github.com/arctanxarc/GENIUS{https://github.com/arctanxarc/GENIUS}.</p> - https://arxiv.org/abs/2602.11144 - Wed, 11 Feb 2026 18:55:54 +0000 - - - Data Repetition Beats Data Scaling in Long-CoT Supervised Fine-Tuning - https://arxiv.org/abs/2602.11149 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.11149.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dawid J. Kopiczko, Sagar Vaze, Tijmen Blankevoort, Yuki M. Asano</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Supervised fine-tuning (SFT) on chain-of-thought data is an essential post-training step for reasoning language models. Standard machine learning intuition suggests that training with more unique training samples yields better generalization. Counterintuitively, we show that SFT benefits from repetition: under a fixed update budget, training for more epochs on smaller datasets outperforms single-epoch training on larger datasets. On AIME'24/25 and GPQA benchmarks, Olmo3-7B trained for 128 epochs on 400 samples outperforms the equivalent 1 epoch on 51200 samples by 12-26 percentage points, with no additional catastrophic forgetting. We find that training token accuracy reliably signals when repetition has saturated; improvements from additional epochs plateau at full memorization, a pattern consistent across all settings. These findings provide a practical approach for reasoning SFT, where scaling epochs with token accuracy as a stopping criterion can replace expensive undirected data scaling. We pose the repetition advantage, where full memorization coincides with improved generalization, as a new open problem for the community in understanding the training dynamics of large language models.</p> - https://arxiv.org/abs/2602.11149 - Wed, 11 Feb 2026 18:58:54 +0000 + Stroke of Surprise: Progressive Semantic Illusions in Vector Sketching + https://arxiv.org/abs/2602.12280 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2602.12280.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Huai-Hsun Cheng, Siang-Ling Zhang, Yu-Lun Liu</p><p><b>Upvotes:</b> 24</p><p><b>Summary:</b> Visual illusions traditionally rely on spatial manipulations such as multi-view consistency. In this work, we introduce Progressive Semantic Illusions, a novel vector sketching task where a single sketch undergoes a dramatic semantic transformation through the sequential addition of strokes. We present Stroke of Surprise, a generative framework that optimizes vector strokes to satisfy distinct semantic interpretations at different drawing stages. The core challenge lies in the "dual-constraint": initial prefix strokes must form a coherent object (e.g., a duck) while simultaneously serving as the structural foundation for a second concept (e.g., a sheep) upon adding delta strokes. To address this, we propose a sequence-aware joint optimization framework driven by a dual-branch Score Distillation Sampling (SDS) mechanism. Unlike sequential approaches that freeze the initial state, our method dynamically adjusts prefix strokes to discover a "common structural subspace" valid for both targets. Furthermore, we introduce a novel Overlay Loss that enforces spatial complementarity, ensuring structural integration rather than occlusion. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art baselines in recognizability and illusion strength, successfully expanding visual anagrams from the spatial to the temporal dimension. Project page: https://stroke-of-surprise.github.io/</p> + https://arxiv.org/abs/2602.12280 + Thu, 12 Feb 2026 18:59:54 +0000