diff --git a/feed.xml b/feed.xml index 4c7403f..56c7b36 100644 --- a/feed.xml +++ b/feed.xml @@ -7,552 +7,412 @@ http://www.rssboard.org/rss-specification python-feedgen en - Wed, 01 Oct 2025 00:02:49 +0000 + Thu, 02 Oct 2025 00:02:38 +0000 - Advancing Reference-free Evaluation of Video Captions with Factual Analysis - https://arxiv.org/abs/2509.16538 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.16538.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shubhashis Roy Dipta, Tz-Ying Wu, Subarna Tripathi</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Video captions offer concise snapshots of actors, objects, and actions within a video, serving as valuable assets for applications such as question answering and event localization. However, acquiring human annotations for video captions is costly or even impractical, especially when dealing with diverse video domains. Existing models trained on supervised datasets face challenges in evaluating performance across different domains due to the reliance on reference-based evaluation protocols, which necessitate ground truth captions. This assumption is unrealistic for evaluating videos in the wild. To address these limitations, we propose a reference-free evaluation framework that does not require ground truth captions, focusing on factual grounding to ensure accurate assessment of caption quality. We introduce VC-Inspector, a novel caption quality evaluator that is both reference-free and factually grounded. Utilizing large language models, we generate pseudo captions of varying quality based on supervised data, which are subsequently used to train a multimodal model (i.e., Qwen2.5-VL) as the evaluator. Our approach demonstrates superior alignment with human judgments on the VATEX-Eval dataset, outperforming existing methods. The performance also generalizes to image caption datasets, Flickr8K-Expert and Flickr8K-CF, when viewing images as 1-frame videos. Overall, VC-Inspector offers a scalable and generalizable solution for evaluating the factual accuracy of video captions, paving the way for more effective and objective assessment methodologies in diverse video domains.</p> - https://arxiv.org/abs/2509.16538 - Sat, 20 Sep 2025 05:04:41 +0000 + Context Is What You Need: The Maximum Effective Context Window for Real World Limits of LLMs + https://arxiv.org/abs/2509.21361 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.21361.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Norman Paulsen</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Large language model (LLM) providers boast big numbers for maximum context window sizes. To test the real world use of context windows, we 1) define a concept of maximum effective context window, 2) formulate a testing method of a context window's effectiveness over various sizes and problem types, and 3) create a standardized way to compare model efficacy for increasingly larger context window sizes to find the point of failure. We collected hundreds of thousands of data points across several models and found significant differences between reported Maximum Context Window (MCW) size and Maximum Effective Context Window (MECW) size. Our findings show that the MECW is, not only, drastically different from the MCW but also shifts based on the problem type. A few top of the line models in our test group failed with as little as 100 tokens in context; most had severe degradation in accuracy by 1000 tokens in context. All models fell far short of their Maximum Context Window by as much as 99 percent. Our data reveals the Maximum Effective Context Window shifts based on the type of problem provided, offering clear and actionable insights into how to improve model accuracy and decrease model hallucination rates.</p> + https://arxiv.org/abs/2509.21361 + Sun, 21 Sep 2025 14:38:17 +0000 - The Photographer Eye: Teaching Multimodal Large Language Models to See and Critique like Photographers - https://arxiv.org/abs/2509.18582 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.18582.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Daiqing Qi, Handong Zhao, Jing Shi, Simon Jenni, Yifei Fan, Franck Dernoncourt, Scott Cohen, Sheng Li</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> While editing directly from life, photographers have found it too difficult to see simultaneously both the blue and the sky. Photographer and curator, Szarkowski insightfully revealed one of the notable gaps between general and aesthetic visual understanding: while the former focuses on identifying the factual element in an image (sky), the latter transcends such object identification, viewing it instead as an aesthetic component--a pure color block (blue). Such fundamental distinctions between general (detection, localization, etc.) and aesthetic (color, lighting, composition, etc.) visual understanding present a significant challenge for Multimodal Large Language Models (MLLMs). Although some recent works have made initial explorations, they are often limited to general and basic aesthetic commonsense. As a result, they frequently fall short in real-world scenarios (Fig. 1), which require extensive expertise--including photographic techniques, photo pre/post-processing knowledge, and more, to provide a detailed analysis and description. To fundamentally enhance the aesthetics understanding of MLLMs, we first introduce a novel dataset, PhotoCritique, derived from extensive discussions among professional photographers and enthusiasts, and characterized by the large scale, expertise, and diversity. Then, to better learn visual aesthetics from PhotoCritique, we furthur propose a novel model, PhotoEye, featuring a languageguided multi-view vision fusion mechanism to understand image aesthetics from multiple perspectives. Finally, we present a novel benchmark, PhotoBench, a comprehensive and professional benchmark for aesthetic visual understanding. On existing benchmarks and PhotoBench, our model demonstrates clear advantages over existing models.</p> - https://arxiv.org/abs/2509.18582 - Tue, 23 Sep 2025 02:59:41 +0000 + GeoRemover: Removing Objects and Their Causal Visual Artifacts + https://arxiv.org/abs/2509.18538 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.18538.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zixin Zhu, Haoxiang Li, Xuelu Feng, He Wu, Chunming Qiao, Junsong Yuan</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Towards intelligent image editing, object removal should eliminate both the target object and its causal visual artifacts, such as shadows and reflections. However, existing image appearance-based methods either follow strictly mask-aligned training and fail to remove these causal effects which are not explicitly masked, or adopt loosely mask-aligned strategies that lack controllability and may unintentionally over-erase other objects. We identify that these limitations stem from ignoring the causal relationship between an object's geometry presence and its visual effects. To address this limitation, we propose a geometry-aware two-stage framework that decouples object removal into (1) geometry removal and (2) appearance rendering. In the first stage, we remove the object directly from the geometry (e.g., depth) using strictly mask-aligned supervision, enabling structure-aware editing with strong geometric constraints. In the second stage, we render a photorealistic RGB image conditioned on the updated geometry, where causal visual effects are considered implicitly as a result of the modified 3D geometry. To guide learning in the geometry removal stage, we introduce a preference-driven objective based on positive and negative sample pairs, encouraging the model to remove objects as well as their causal visual artifacts while avoiding new structural insertions. Extensive experiments demonstrate that our method achieves state-of-the-art performance in removing both objects and their associated artifacts on two popular benchmarks. The code is available at https://github.com/buxiangzhiren/GeoRemover.</p> + https://arxiv.org/abs/2509.18538 + Tue, 23 Sep 2025 02:04:19 +0000 - Charting a Decade of Computational Linguistics in Italy: The CLiC-it Corpus - https://arxiv.org/abs/2509.19033 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.19033.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chiara Alzetta, Serena Auriemma, Alessandro Bondielli, Luca Dini, Chiara Fazzone, Alessio Miaschi, Martina Miliani, Marta Sartor</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Over the past decade, Computational Linguistics (CL) and Natural Language Processing (NLP) have evolved rapidly, especially with the advent of Transformer-based Large Language Models (LLMs). This shift has transformed research goals and priorities, from Lexical and Semantic Resources to Language Modelling and Multimodality. In this study, we track the research trends of the Italian CL and NLP community through an analysis of the contributions to CLiC-it, arguably the leading Italian conference in the field. We compile the proceedings from the first 10 editions of the CLiC-it conference (from 2014 to 2024) into the CLiC-it Corpus, providing a comprehensive analysis of both its metadata, including author provenance, gender, affiliations, and more, as well as the content of the papers themselves, which address various topics. Our goal is to provide the Italian and international research communities with valuable insights into emerging trends and key developments over time, supporting informed decisions and future directions in the field.</p> - https://arxiv.org/abs/2509.19033 - Tue, 23 Sep 2025 14:06:09 +0000 + Benefits and Pitfalls of Reinforcement Learning for Language Model Planning: A Theoretical Perspective + https://arxiv.org/abs/2509.22613 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22613.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Siwei Wang, Yifei Shen, Haoran Sun, Shi Feng, Shang-Hua Teng, Li Dong, Yaru Hao, Wei Chen</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Recent reinforcement learning (RL) methods have substantially enhanced the planning capabilities of Large Language Models (LLMs), yet the theoretical basis for their effectiveness remains elusive. In this work, we investigate RL's benefits and limitations through a tractable graph-based abstraction, focusing on policy gradient (PG) and Q-learning methods. Our theoretical analyses reveal that supervised fine-tuning (SFT) may introduce co-occurrence-based spurious solutions, whereas RL achieves correct planning primarily through exploration, underscoring exploration's role in enabling better generalization. However, we also show that PG suffers from diversity collapse, where output diversity decreases during training and persists even after perfect accuracy is attained. By contrast, Q-learning provides two key advantages: off-policy learning and diversity preservation at convergence. We further demonstrate that careful reward design is necessary to prevent reward hacking in Q-learning. Finally, applying our framework to the real-world planning benchmark Blocksworld, we confirm that these behaviors manifest in practice.</p> + https://arxiv.org/abs/2509.22613 + Fri, 26 Sep 2025 17:39:48 +0000 - Combinatorial Creativity: A New Frontier in Generalization Abilities - https://arxiv.org/abs/2509.21043 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.21043.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Samuel Schapiro, Sumuk Shashidhar, Alexi Gladstone, Jonah Black, Royce Moon, Dilek Hakkani-Tur, Lav R. Varshney</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Artificial intelligence (AI) systems, and Large Language Models (LLMs) in particular, are increasingly employed for creative tasks like scientific idea generation, constituting a form of generalization from training data unaddressed by existing conceptual frameworks. Despite its similarities to compositional generalization (CG), combinatorial creativity (CC) is an open-ended ability. Instead of evaluating for accuracy or correctness against fixed targets, which would contradict the open-ended nature of CC, we propose a theoretical framework and algorithmic task for evaluating outputs by their degrees of novelty and utility. From here, we make several important empirical contributions: (1) We obtain the first insights into the scaling behavior of creativity for LLMs. (2) We discover that, for fixed compute budgets, there exist optimal model depths and widths for creative ability. (3) We find that the ideation-execution gap, whereby LLMs excel at generating novel scientific ideas but struggle to ensure their practical feasibility, may be explained by a more fundamental novelty-utility tradeoff characteristic of creativity algorithms in general. Importantly, this tradeoff remains persistent even at scale, casting doubt on the long-term creative potential of LLMs in their current form. Together, our conceptual framework and empirical findings provide a foundation for understanding and improving creativity in modern AI models, bridging the gap between human and machine intelligence.</p> - https://arxiv.org/abs/2509.21043 - Thu, 25 Sep 2025 11:48:37 +0000 + Learning Human-Perceived Fakeness in AI-Generated Videos via Multimodal LLMs + https://arxiv.org/abs/2509.22646 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22646.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xingyu Fu, Siyi Liu, Yinuo Xu, Pan Lu, Guangqiuse Hu, Tianbo Yang, Taran Anantasagar, Christopher Shen, Yikai Mao, Yuanzhe Liu, Keyush Shah, Chung Un Lee, Yejin Choi, James Zou, Dan Roth, Chris Callison-Burch</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Can humans identify AI-generated (fake) videos and provide grounded reasons? While video generation models have advanced rapidly, a critical dimension -- whether humans can detect deepfake traces within a generated video, i.e., spatiotemporal grounded visual artifacts that reveal a video as machine generated -- has been largely overlooked. We introduce DeeptraceReward, the first fine-grained, spatially- and temporally- aware benchmark that annotates human-perceived fake traces for video generation reward. The dataset comprises 4.3K detailed annotations across 3.3K high-quality generated videos. Each annotation provides a natural-language explanation, pinpoints a bounding-box region containing the perceived trace, and marks precise onset and offset timestamps. We consolidate these annotations into 9 major categories of deepfake traces that lead humans to identify a video as AI-generated, and train multimodal language models (LMs) as reward models to mimic human judgments and localizations. On DeeptraceReward, our 7B reward model outperforms GPT-5 by 34.7% on average across fake clue identification, grounding, and explanation. Interestingly, we observe a consistent difficulty gradient: binary fake v.s. real classification is substantially easier than fine-grained deepfake trace detection; within the latter, performance degrades from natural language explanations (easiest), to spatial grounding, to temporal labeling (hardest). By foregrounding human-perceived deepfake traces, DeeptraceReward provides a rigorous testbed and training signal for socially aware and trustworthy video generation.</p> + https://arxiv.org/abs/2509.22646 + Fri, 26 Sep 2025 17:59:54 +0000 - LUMINA: Detecting Hallucinations in RAG System with Context-Knowledge Signals - https://arxiv.org/abs/2509.21875 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.21875.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Min-Hsuan Yeh, Yixuan Li, Tanwi Mallick</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Retrieval-Augmented Generation (RAG) aims to mitigate hallucinations in large language models (LLMs) by grounding responses in retrieved documents. Yet, RAG-based LLMs still hallucinate even when provided with correct and sufficient context. A growing line of work suggests that this stems from an imbalance between how models use external context and their internal knowledge, and several approaches have attempted to quantify these signals for hallucination detection. However, existing methods require extensive hyperparameter tuning, limiting their generalizability. We propose LUMINA, a novel framework that detects hallucinations in RAG systems through context-knowledge signals: external context utilization is quantified via distributional distance, while internal knowledge utilization is measured by tracking how predicted tokens evolve across transformer layers. We further introduce a framework for statistically validating these measurements. Experiments on common RAG hallucination benchmarks and four open-source LLMs show that LUMINA achieves consistently high AUROC and AUPRC scores, outperforming prior utilization-based methods by up to +13% AUROC on HalluRAG. Moreover, LUMINA remains robust under relaxed assumptions about retrieval quality and model matching, offering both effectiveness and practicality.</p> - https://arxiv.org/abs/2509.21875 - Fri, 26 Sep 2025 04:57:46 +0000 + Convolutional Set Transformer + https://arxiv.org/abs/2509.22889 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22889.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Federico Chinello, Giacomo Boracchi</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> We introduce the Convolutional Set Transformer (CST), a novel neural architecture designed to process image sets of arbitrary cardinality that are visually heterogeneous yet share high-level semantics - such as a common category, scene, or concept. Existing set-input networks, e.g., Deep Sets and Set Transformer, are limited to vector inputs and cannot directly handle 3D image tensors. As a result, they must be cascaded with a feature extractor, typically a CNN, which encodes images into embeddings before the set-input network can model inter-image relationships. In contrast, CST operates directly on 3D image tensors, performing feature extraction and contextual modeling simultaneously, thereby enabling synergies between the two processes. This design yields superior performance in tasks such as Set Classification and Set Anomaly Detection and further provides native compatibility with CNN explainability methods such as Grad-CAM, unlike competing approaches that remain opaque. Finally, we show that CSTs can be pre-trained on large-scale datasets and subsequently adapted to new domains and tasks through standard Transfer Learning schemes. To support further research, we release CST-15, a CST backbone pre-trained on ImageNet (https://github.com/chinefed/convolutional-set-transformer).</p> + https://arxiv.org/abs/2509.22889 + Fri, 26 Sep 2025 20:13:00 +0000 - MultiCrafter: High-Fidelity Multi-Subject Generation via Spatially Disentangled Attention and Identity-Aware Reinforcement Learning - https://arxiv.org/abs/2509.21953 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.21953.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tao Wu, Yibo Jiang, Yehao Lu, Zhizhong Wang, Zeyi Huang, Zequn Qin, Xi Li</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Multi-subject image generation aims to synthesize user-provided subjects in a single image while preserving subject fidelity, ensuring prompt consistency, and aligning with human aesthetic preferences. However, existing methods, particularly those built on the In-Context-Learning paradigm, are limited by their reliance on simple reconstruction-based objectives, leading to both severe attribute leakage that compromises subject fidelity and failing to align with nuanced human preferences. To address this, we propose MultiCrafter, a framework that ensures high-fidelity, preference-aligned generation. First, we find that the root cause of attribute leakage is a significant entanglement of attention between different subjects during the generation process. Therefore, we introduce explicit positional supervision to explicitly separate attention regions for each subject, effectively mitigating attribute leakage. To enable the model to accurately plan the attention region of different subjects in diverse scenarios, we employ a Mixture-of-Experts architecture to enhance the model's capacity, allowing different experts to focus on different scenarios. Finally, we design a novel online reinforcement learning framework to align the model with human preferences, featuring a scoring mechanism to accurately assess multi-subject fidelity and a more stable training strategy tailored for the MoE architecture. Experiments validate that our framework significantly improves subject fidelity while aligning with human preferences better.</p> - https://arxiv.org/abs/2509.21953 - Fri, 26 Sep 2025 06:41:43 +0000 + LLM Watermark Evasion via Bias Inversion + https://arxiv.org/abs/2509.23019 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23019.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jeongyeon Hwang, Sangdon Park, Jungseul Ok</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Watermarking for large language models (LLMs) embeds a statistical signal during generation to enable detection of model-produced text. While watermarking has proven effective in benign settings, its robustness under adversarial evasion remains contested. To advance a rigorous understanding and evaluation of such vulnerabilities, we propose the Bias-Inversion Rewriting Attack (BIRA), which is theoretically motivated and model-agnostic. BIRA weakens the watermark signal by suppressing the logits of likely watermarked tokens during LLM-based rewriting, without any knowledge of the underlying watermarking scheme. Across recent watermarking methods, BIRA achieves over 99\% evasion while preserving the semantic content of the original text. Beyond demonstrating an attack, our results reveal a systematic vulnerability, emphasizing the need for stress testing and robust defenses.</p> + https://arxiv.org/abs/2509.23019 + Sat, 27 Sep 2025 00:24:57 +0000 - When Does Reasoning Matter? A Controlled Study of Reasoning's Contribution to Model Performance - https://arxiv.org/abs/2509.22193 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22193.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Nicolas Boizard, Hippolyte Gisserot-Boukhlef, Kevin El-Haddad, Céline Hudelot, Pierre Colombo</p><p><b>Upvotes:</b> 29</p><p><b>Summary:</b> Large Language Models (LLMs) with reasoning capabilities have achieved state-of-the-art performance on a wide range of tasks. Despite its empirical success, the tasks and model scales at which reasoning becomes effective, as well as its training and inference costs, remain underexplored. In this work, we rely on a synthetic data distillation framework to conduct a large-scale supervised study. We compare Instruction Fine-Tuning (IFT) and reasoning models of varying sizes, on a wide range of math-centric and general-purpose tasks, evaluating both multiple-choice and open-ended formats. Our analysis reveals that reasoning consistently improves model performance, often matching or surpassing significantly larger IFT systems. Notably, while IFT remains Pareto-optimal in training and inference costs, reasoning models become increasingly valuable as model size scales, overcoming IFT performance limits on reasoning-intensive and open-ended tasks.</p> - https://arxiv.org/abs/2509.22193 - Fri, 26 Sep 2025 10:53:52 +0000 + BuildBench: Benchmarking LLM Agents on Compiling Real-World Open-Source Software + https://arxiv.org/abs/2509.25248 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25248.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zehua Zhang, Ati Priya Bajaj, Divij Handa, Siyu Liu, Arvind S Raj, Hongkai Chen, Hulin Wang, Yibo Liu, Zion Leonahenahe Basque, Souradip Nath, Vishal Juneja, Nikhil Chapre, Yan Shoshitaishvili, Adam Doupé, Chitta Baral, Ruoyu Wang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Automatically compiling open-source software (OSS) projects is a vital, labor-intensive, and complex task, which makes it a good challenge for LLM Agents. Existing methods rely on manually curated rules and workflows, which cannot adapt to OSS that requires customized configuration or environment setup. Recent attempts using Large Language Models (LLMs) used selective evaluation on a subset of highly rated OSS, a practice that underestimates the realistic challenges of OSS compilation. In practice, compilation instructions are often absent, dependencies are undocumented, and successful builds may even require patching source files or modifying build scripts. We propose a more challenging and realistic benchmark, BUILD-BENCH, comprising OSS that are more diverse in quality, scale, and characteristics. Furthermore, we propose a strong baseline LLM-based agent, OSS-BUILD-AGENT, an effective system with enhanced build instruction retrieval module that achieves state-of-the-art performance on BUILD-BENCH and is adaptable to heterogeneous OSS characteristics. We also provide detailed analysis regarding different compilation method design choices and their influence to the whole task, offering insights to guide future advances. We believe performance on BUILD-BENCH can faithfully reflect an agent's ability to tackle compilation as a complex software engineering tasks, and, as such, our benchmark will spur innovation with a significant impact on downstream applications in the fields of software development and software security.</p> + https://arxiv.org/abs/2509.25248 + Sat, 27 Sep 2025 03:02:46 +0000 - StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs - https://arxiv.org/abs/2509.22220 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22220.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuhan Song, Linhao Zhang, Chuhan Wu, Aiwei Liu, Wei Jia, Houfeng Wang, Xiao Zhou</p><p><b>Upvotes:</b> 57</p><p><b>Summary:</b> Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks.</p> - https://arxiv.org/abs/2509.22220 - Fri, 26 Sep 2025 11:32:51 +0000 + d^2Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching + https://arxiv.org/abs/2509.23094 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23094.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuchu Jiang, Yue Cai, Xiangzhong Luo, Jiale Fu, Jiarui Wang, Chonghan Liu, Xu Yang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Diffusion-based large language models (dLLMs), despite their promising performance, still suffer from inferior inference efficiency. This is because dLLMs rely on bidirectional attention and cannot directly benefit from the standard key-value (KV) cache as autoregressive models (ARMs) do. To tackle this issue, we introduce Dual aDaptive Cache (d^2Cache), which is a training-free approximate KV cache framework for accelerating dLLM inference. d^2Cache features a two-stage fine-grained selection strategy to identify tokens and adaptively update their KV states at each decoding step, while caching the KV states of the remaining tokens for reuse. Furthermore, d^2Cache naturally offers a more reliable decoding alternative, which can enable quasi left-to-right generation and mitigate premature overconfidence in tokens at the end of the sequence. Extensive experimental results on two representative dLLMs (\ie, LLaDA and Dream) demonstrate that d^2Cache not only achieves substantial inference speedups, but also yields consistent improvements in generation quality. The code is available at https://github.com/Kamichanw/d2Cache.</p> + https://arxiv.org/abs/2509.23094 + Sat, 27 Sep 2025 04:07:23 +0000 - REMA: A Unified Reasoning Manifold Framework for Interpreting Large Language Model - https://arxiv.org/abs/2509.22518 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22518.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Bo Li, Guanzhi Deng, Ronghao Chen, Junrong Yue, Shuo Zhang, Qinghua Zhao, Linqi Song, Lijie Wen</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Understanding how Large Language Models (LLMs) perform complex reasoning and their failure mechanisms is a challenge in interpretability research. To provide a measurable geometric analysis perspective, we define the concept of the Reasoning Manifold, a latent low-dimensional geometric structure formed by the internal representations corresponding to all correctly reasoned generations. This structure can be conceptualized as the embodiment of the effective thinking paths that the model has learned to successfully solve a given task. Based on this concept, we build REMA, a framework that explains the origins of failures by quantitatively comparing the spatial relationships of internal model representations corresponding to both erroneous and correct reasoning samples. Specifically, REMA first quantifies the geometric deviation of each erroneous representation by calculating its k-nearest neighbors distance to the approximated manifold formed by correct representations, thereby providing a unified failure signal. It then localizes the divergence points where these deviations first become significant by tracking this deviation metric across the model's layers and comparing it against a baseline of internal fluctuations from correct representations, thus identifying where the reasoning chain begins to go off-track. Our extensive experiments on diverse language and multimodal models and tasks demonstrate the low-dimensional nature of the reasoning manifold and the high separability between erroneous and correct reasoning representations. The results also validate the effectiveness of the REMA framework in analyzing the origins of reasoning failures. This research connects abstract reasoning failures to measurable geometric deviations in representations, providing new avenues for in-depth understanding and diagnosis of the internal computational processes of black-box models.</p> - https://arxiv.org/abs/2509.22518 - Fri, 26 Sep 2025 16:02:27 +0000 + Test-Time Policy Adaptation for Enhanced Multi-Turn Interactions with LLMs + https://arxiv.org/abs/2509.23166 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23166.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chenxing Wei, Hong Wang, Ying He, Fei Yu, Yao Shu</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Large Language Models (LLMs) employ multi-turn interaction as a fundamental paradigm for completing complex tasks. However, their performance often degrades in extended interactions, as they are typically trained on static, single-turn data, which hinders their ability to adapt to real-time user feedback. To address this limitation, we first propose a new paradigm: Test-Time Policy Adaptation for Multi-Turn Interactions (T2PAM), which utilizes user feedback from the ongoing interaction as a reward signal to estimate a latent optimal policy aligned with user preferences, then updates a small subset of parameters to steer the model toward this policy, ultimately enabling efficient in-conversation self-correction. We then introduce Optimum-Referenced One-Step Adaptation (ROSA), a lightweight algorithm that operationalizes T2PAM. ROSA guides the model parameters toward a theoretical optimal policy in a single, efficient update step, avoiding costly iterative gradient-based optimization and minimizing computational overhead. We provide a rigorous theoretical analysis guaranteeing that the policy of ROSA converges to the preference of user as the number of interactions increases. Extensive experiments on challenging benchmark demonstrate that ROSA achieves significant improvements in both task effectiveness and efficiency.</p> + https://arxiv.org/abs/2509.23166 + Sat, 27 Sep 2025 07:46:15 +0000 - UniMIC: Token-Based Multimodal Interactive Coding for Human-AI Collaboration - https://arxiv.org/abs/2509.22570 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22570.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qi Mao, Tinghan Yang, Jiahao Li, Bin Li, Libiao Jin, Yan Lu</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> The rapid progress of Large Multimodal Models (LMMs) and cloud-based AI agents is transforming human-AI collaboration into bidirectional, multimodal interaction. However, existing codecs remain optimized for unimodal, one-way communication, resulting in repeated degradation under conventional compress-transmit-reconstruct pipelines. To address this limitation, we propose UniMIC, a Unified token-based Multimodal Interactive Coding framework that bridges edge devices and cloud AI agents. Instead of transmitting raw pixels or plain text, UniMIC employs compact tokenized representations as the communication medium, enabling efficient low-bitrate transmission while maintaining compatibility with LMMs. To further enhance compression, lightweight Transformer-based entropy models with scenario-specific designs-generic, masked, and text-conditioned-effectively minimize inter-token redundancy. Extensive experiments on text-to-image generation, text-guided inpainting, outpainting, and visual question answering show that UniMIC achieves substantial bitrate savings and remains robust even at ultra-low bitrates (<0.05bpp), without compromising downstream task performance. These results establish UniMIC as a practical and forward-looking paradigm for next-generation multimodal interactive communication.</p> - https://arxiv.org/abs/2509.22570 - Fri, 26 Sep 2025 16:46:12 +0000 + Efficient Audio-Visual Speech Separation with Discrete Lip Semantics and Multi-Scale Global-Local Attention + https://arxiv.org/abs/2509.23610 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23610.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kai Li, Kejun Gao, Xiaolin Hu</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Audio-visual speech separation (AVSS) methods leverage visual cues to extract target speech and have demonstrated strong separation quality in noisy acoustic environments. However, these methods usually involve a large number of parameters and require high computational cost, which is unacceptable in many applications where speech separation serves as only a preprocessing step for further speech processing. To address this issue, we propose an efficient AVSS method, named Dolphin. For visual feature extraction, we develop DP-LipCoder, a dual-path lightweight video encoder that transforms lip-motion into discrete audio-aligned semantic tokens. For audio separation, we construct a lightweight encoder-decoder separator, in which each layer incorporates a global-local attention (GLA) block to efficiently capture multi-scale dependencies. Experiments on three benchmark datasets showed that Dolphin not only surpassed the current state-of-the-art (SOTA) model in separation quality but also achieved remarkable improvements in efficiency: over 50% fewer parameters, more than 2.4x reduction in MACs, and over 6x faster GPU inference speed. These results indicate that Dolphin offers a practical and deployable solution for high-performance AVSS in real-world scenarios. Our code and demo page are publicly available at http://cslikai.cn/Dolphin/.</p> + https://arxiv.org/abs/2509.23610 + Sun, 28 Sep 2025 03:25:34 +0000 - Dynamic Experts Search: Enhancing Reasoning in Mixture-of-Experts LLMs at Test Time - https://arxiv.org/abs/2509.22572 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22572.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yixuan Han, Fan Ma, Ruijie Quan, Yi Yang</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Test-Time Scaling (TTS) enhances the reasoning ability of large language models (LLMs) by allocating additional computation during inference. However, existing approaches primarily rely on output-level sampling while overlooking the role of model architecture. In mainstream Mixture-of-Experts (MoE) LLMs, we observe that varying the number of activated experts yields complementary solution sets with stable accuracy, revealing a new and underexplored source of diversity. Motivated by this observation, we propose Dynamic Experts Search (DES), a TTS strategy that elevates expert activation into a controllable dimension of the search space. DES integrates two key components: (1) Dynamic MoE, which enables direct control of expert counts during inference to generate diverse reasoning trajectories without additional cost; and (2) Expert Configuration Inheritance, which preserves consistent expert counts within a reasoning path while varying them across runs, thereby balancing stability and diversity throughout the search. Extensive experiments across MoE architectures, verifiers and reasoning benchmarks (i.e., math, code and knowledge) demonstrate that DES reliably outperforms TTS baselines, enhancing accuracy and stability without additional cost. These results highlight DES as a practical and scalable form of architecture-aware TTS, illustrating how structural flexibility in modern LLMs can advance reasoning.</p> - https://arxiv.org/abs/2509.22572 - Fri, 26 Sep 2025 16:49:10 +0000 + Estimating Time Series Foundation Model Transferability via In-Context Learning + https://arxiv.org/abs/2509.23695 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23695.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qingren Yao, Ming Jin, Chengqi Zhang, Chao-Han Huck Yang, Jun Qi, Shirui Pan</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Time series foundation models (TSFMs) offer strong zero-shot forecasting via large-scale pre-training, yet fine-tuning remains critical for boosting performance in domains with limited public data. With the growing number of TSFMs, efficiently identifying the best model for downstream fine-tuning becomes increasingly challenging. In this work, we introduce TimeTic, a transferability estimation framework that recasts model selection as an in-context-learning problem: given observations on known (source) datasets, it predicts how a TSFM will perform after fine-tuning on a downstream (target) dataset. TimeTic flexibly organizes the observed model-data relationships as contextual information, allowing it to adapt seamlessly to various test-time scenarios. Leveraging the natural tabular structure formed by dataset meta-features, model characteristics, and fine-tuned performance, we employ tabular foundation models to serve as in-context learners. We further introduce a novel model characterization based on entropy evolution across model layers, capturing embedding-space distinctions and enabling TimeTic to generalize across arbitrary model sets. We establish a comprehensive benchmark for transferability estimation including 10 datasets, 10 foundation models, and 3 forecasting tasks. On this benchmark, TimeTic's estimation demonstrates strong alignment with actual fine-tuned performance for previously unseen datasets, achieving a mean rank correlation of approximately 0.6 and a 30% improvement compared to using zero-shot performance as the transferability score.</p> + https://arxiv.org/abs/2509.23695 + Sun, 28 Sep 2025 07:07:13 +0000 - VideoScore2: Think before You Score in Generative Video Evaluation - https://arxiv.org/abs/2509.22799 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22799.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xuan He, Dongfu Jiang, Ping Nie, Minghao Liu, Zhengxuan Jiang, Mingyi Su, Wentao Ma, Junru Lin, Chun Ye, Yi Lu, Keming Wu, Benjamin Schneider, Quy Duc Do, Zhuofeng Li, Yiming Jia, Yuxuan Zhang, Guo Cheng, Haozhe Wang, Wangchunshu Zhou, Qunshu Lin, Yuanxing Zhang, Ge Zhang, Wenhao Huang, Wenhu Chen</p><p><b>Upvotes:</b> 20</p><p><b>Summary:</b> Recent advances in text-to-video generation have produced increasingly realistic and diverse content, yet evaluating such videos remains a fundamental challenge due to their multi-faceted nature encompassing visual quality, semantic alignment, and physical consistency. Existing evaluators and reward models are limited to single opaque scores, lack interpretability, or provide only coarse analysis, making them insufficient for capturing the comprehensive nature of video quality assessment. We present VideoScore2, a multi-dimensional, interpretable, and human-aligned framework that explicitly evaluates visual quality, text-to-video alignment, and physical/common-sense consistency while producing detailed chain-of-thought rationales. Our model is trained on a large-scale dataset VideoFeedback2 containing 27,168 human-annotated videos with both scores and reasoning traces across three dimensions, using a two-stage pipeline of supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization (GRPO) to enhance analytical robustness. Extensive experiments demonstrate that VideoScore2 achieves superior performance with 44.35 (+5.94) accuracy on our in-domain benchmark VideoScore-Bench-v2 and 50.37 (+4.32) average performance across four out-of-domain benchmarks (VideoGenReward-Bench, VideoPhy2, etc), while providing interpretable assessments that bridge the gap between evaluation and controllable generation through effective reward modeling for Best-of-N sampling. Project Page: https://tiger-ai-lab.github.io/VideoScore2/</p> - https://arxiv.org/abs/2509.22799 - Fri, 26 Sep 2025 18:09:03 +0000 + Knowledge Homophily in Large Language Models + https://arxiv.org/abs/2509.23773 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23773.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Utkarsh Sahu, Zhisheng Qi, Mahantesh Halappanavar, Nedim Lipka, Ryan A. Rossi, Franck Dernoncourt, Yu Zhang, Yao Ma, Yu Wang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Large Language Models (LLMs) have been increasingly studied as neural knowledge bases for supporting knowledge-intensive applications such as question answering and fact checking. However, the structural organization of their knowledge remains unexplored. Inspired by cognitive neuroscience findings, such as semantic clustering and priming, where knowing one fact increases the likelihood of recalling related facts, we investigate an analogous knowledge homophily pattern in LLMs. To this end, we map LLM knowledge into a graph representation through knowledge checking at both the triplet and entity levels. After that, we analyze the knowledgeability relationship between an entity and its neighbors, discovering that LLMs tend to possess a similar level of knowledge about entities positioned closer in the graph. Motivated by this homophily principle, we propose a Graph Neural Network (GNN) regression model to estimate entity-level knowledgeability scores for triplets by leveraging their neighborhood scores. The predicted knowledgeability enables us to prioritize checking less well-known triplets, thereby maximizing knowledge coverage under the same labeling budget. This not only improves the efficiency of active labeling for fine-tuning to inject knowledge into LLMs but also enhances multi-hop path retrieval in reasoning-intensive question answering.</p> + https://arxiv.org/abs/2509.23773 + Sun, 28 Sep 2025 09:40:27 +0000 - MMPB: It's Time for Multi-Modal Personalization - https://arxiv.org/abs/2509.22820 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22820.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jaeik Kim, Woojin Kim, Woohyeon Park, Jaeyoung Do</p><p><b>Upvotes:</b> 14</p><p><b>Summary:</b> Visual personalization is essential in user-facing AI systems such as smart homes and healthcare, where aligning model behavior with user-centric concepts is critical. However, recent large Vision-Language Models (VLMs), despite their broad applicability, remain underexplored in their ability to adapt to individual users. In this paper, we introduce MMPB, the first extensive benchmark for evaluating VLMs on personalization. MMPB comprises 10k image-query pairs and includes 111 personalizable concepts across four categories: humans, animals, objects, and characters, with the human category enriched with preference-grounded queries. We structure personalization into three main task types, each highlighting a different key property of VLMs. Using 23 widely used VLMs including both open- and closed-source models, we evaluate personalization performance via a three-stage protocol: concept injection, multi-turn dialogue, and personalized querying. Our findings indicate that most VLMs (including some closed-source models) struggle with personalization, particularly in maintaining consistency over dialogue, handling user preferences, and adapting to visual cues. Our analysis reveals that the challenges in VLM personalization (such as refusal behaviors and long-context forgetting) highlight substantial room for improvement. By identifying these limitations and offering a scalable benchmark, MMPB offers valuable insights and a solid foundation for future research toward truly personalized multi-modal AI. Project Page: aidaslab.github.io/MMPB</p> - https://arxiv.org/abs/2509.22820 - Fri, 26 Sep 2025 18:24:48 +0000 + Winning the Pruning Gamble: A Unified Approach to Joint Sample and Token Pruning for Efficient Supervised Fine-Tuning + https://arxiv.org/abs/2509.23873 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23873.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shaobo Wang, Jiaming Wang, Jiajun Zhang, Cong Wang, Yue Min, Zichen Wen, Fei Huang, Huiqiang Jiang, Junyang Lin, Dayiheng Liu, Linfeng Zhang</p><p><b>Upvotes:</b> 55</p><p><b>Summary:</b> As supervised fine-tuning (SFT) evolves from a lightweight post-training step into a compute-intensive phase rivaling mid-training in scale, data efficiency has become critical for aligning large language models (LLMs) under tight budgets. Existing data pruning methods suffer from a fragmented design: they operate either at the sample level or the token level in isolation, failing to jointly optimize both dimensions. This disconnect leads to significant inefficiencies--high-value samples may still contain redundant tokens, while token-level pruning often discards crucial instructional or corrective signals embedded in individual examples. To address this bottleneck, we introduce the Error-Uncertainty (EU) Plane, a diagnostic framework that jointly characterizes the heterogeneous utility of training data across samples and tokens. Guided by this insight, we propose Quadrant-based Tuning (Q-Tuning), a unified framework that strategically coordinates sample pruning and token pruning. Q-Tuning employs a two-stage strategy: first, it performs sample-level triage to retain examples rich in informative misconceptions or calibration signals; second, it applies an asymmetric token-pruning policy, using a context-aware scoring mechanism to trim less salient tokens exclusively from misconception samples while preserving calibration samples in their entirety. Our method sets a new state of the art across five diverse benchmarks. Remarkably, on SmolLM2-1.7B, Q-Tuning achieves a +38\% average improvement over the full-data SFT baseline using only 12.5\% of the original training data. As the first dynamic pruning approach to consistently outperform full-data training, Q-Tuning provides a practical and scalable blueprint for maximizing data utilization in budget-constrained LLM SFT.</p> + https://arxiv.org/abs/2509.23873 + Sun, 28 Sep 2025 13:27:38 +0000 - Critique-Coder: Enhancing Coder Models by Critique Reinforcement Learning - https://arxiv.org/abs/2509.22824 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22824.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chi Ruan, Dongfu Jiang, Yubo Wang, Wenhu Chen</p><p><b>Upvotes:</b> 18</p><p><b>Summary:</b> Reinforcement Learning (RL) has emerged as a popular training paradigm, particularly when paired with reasoning models. While effective, it primarily focuses on generating responses and lacks mechanisms to explicitly foster critique or reflection. Several recent studies, like Critique-Fine-Tuning (CFT) and Critique-Guided-Distillation (CGD) have shown the benefits of explicitly teaching LLMs how to critique. Motivated by them, we propose Critique Reinforcement Learning (CRL), where the model is tasked with generating a critique for a given (question, solution) pair. The reward is determined solely by whether the final judgment label c in {True, False} of the generated critique aligns with the ground-truth judgment c^*. Building on this point, we introduce Critique-Coder, which is trained on a hybrid of RL and CRL by substituting 20\% of the standard RL data with CRL data. We fine-tune multiple models (Critique-Coder) and evaluate them on different benchmarks to show their advantages over RL-only models. We show that Critique-Coder consistently outperforms RL-only baselines on all the evaluated benchmarks. Notably, our Critique-Coder-8B can reach over 60\% on LiveCodeBench (v5), outperforming other reasoning models like DeepCoder-14B and GPT-o1. Beyond code generation, Critique-Coder also demonstrates enhanced general reasoning abilities, as evidenced by its better performance on logic reasoning tasks from the BBEH dataset. This indicates that the application of CRL on coding datasets enhances general reasoning and critique abilities, which are transferable across a broad range of tasks. Hence, we believe that CRL works as a great complement to standard RL for LLM reasoning.</p> - https://arxiv.org/abs/2509.22824 - Fri, 26 Sep 2025 18:30:49 +0000 + MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use + https://arxiv.org/abs/2509.24002 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24002.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zijian Wu, Xiangyan Liu, Xinyuan Zhang, Lingjun Chen, Fanqing Meng, Lingxiao Du, Yiran Zhao, Fanshi Zhang, Yaoqi Ye, Jiawei Wang, Zirui Wang, Jinjie Ni, Yufan Yang, Arvin Xu, Michael Qizhe Shieh</p><p><b>Upvotes:</b> 119</p><p><b>Summary:</b> MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of 127 high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only 52.56\% pass@1 and 33.86\% pass^4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below 30\% pass@1 and 15\% pass^4. On average, LLMs require 16.2 execution turns and 17.4 tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.</p> + https://arxiv.org/abs/2509.24002 + Sun, 28 Sep 2025 17:53:27 +0000 - ChatInject: Abusing Chat Templates for Prompt Injection in LLM Agents - https://arxiv.org/abs/2509.22830 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22830.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hwan Chang, Yonghyun Jun, Hwanhee Lee</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> The growing deployment of large language model (LLM) based agents that interact with external environments has created new attack surfaces for adversarial manipulation. One major threat is indirect prompt injection, where attackers embed malicious instructions in external environment output, causing agents to interpret and execute them as if they were legitimate prompts. While previous research has focused primarily on plain-text injection attacks, we find a significant yet underexplored vulnerability: LLMs' dependence on structured chat templates and their susceptibility to contextual manipulation through persuasive multi-turn dialogues. To this end, we introduce ChatInject, an attack that formats malicious payloads to mimic native chat templates, thereby exploiting the model's inherent instruction-following tendencies. Building on this foundation, we develop a persuasion-driven Multi-turn variant that primes the agent across conversational turns to accept and execute otherwise suspicious actions. Through comprehensive experiments across frontier LLMs, we demonstrate three critical findings: (1) ChatInject achieves significantly higher average attack success rates than traditional prompt injection methods, improving from 5.18% to 32.05% on AgentDojo and from 15.13% to 45.90% on InjecAgent, with multi-turn dialogues showing particularly strong performance at average 52.33% success rate on InjecAgent, (2) chat-template-based payloads demonstrate strong transferability across models and remain effective even against closed-source LLMs, despite their unknown template structures, and (3) existing prompt-based defenses are largely ineffective against this attack approach, especially against Multi-turn variants. These findings highlight vulnerabilities in current agent systems.</p> - https://arxiv.org/abs/2509.22830 - Fri, 26 Sep 2025 18:38:07 +0000 + CORRECT: COndensed eRror RECognition via knowledge Transfer in multi-agent systems + https://arxiv.org/abs/2509.24088 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24088.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yifan Yu, Moyan Li, Shaoyuan Xu, Jinmiao Fu, Xinhai Hou, Fan Lai, Bryan Wang</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Multi-agent systems (MAS) are increasingly capable of tackling complex real-world tasks, yet their reliance on inter-agent coordination, tool use, and long-horizon reasoning makes error recognition particularly challenging. Minor errors can propagate across agents, escalating into task failures while producing long, intertwined execution trajectories that impose significant costs for both human developers and automated systems to debug and analyze. Our key insight is that, despite surface differences in failure trajectories (e.g., logs), MAS errors often recur with similar structural patterns. This paper presents CORRECT, the first lightweight, training-free framework that leverages an online cache of distilled error schemata to recognize and transfer knowledge of failure structures across new requests. This cache-based reuse allows LLMs to perform targeted error localization at inference time, avoiding the need for expensive retraining while adapting to dynamic MAS deployments in subseconds. To support rigorous study in this domain, we also introduce CORRECT-Error, a large-scale dataset of over 2,000 annotated trajectories collected through a novel error-injection pipeline guided by real-world distributions, and further validated through human evaluation to ensure alignment with natural failure patterns. Experiments across seven diverse MAS applications show that CORRECT improves step-level error localization up to 19.8% over existing advances while at near-zero overhead, substantially narrowing the gap between automated and human-level error recognition.</p> + https://arxiv.org/abs/2509.24088 + Sun, 28 Sep 2025 21:47:20 +0000 - Rethinking Large Language Model Distillation: A Constrained Markov Decision Process Perspective - https://arxiv.org/abs/2509.22921 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22921.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Matthieu Zimmer, Xiaotong Ji, Tu Nguyen, Haitham Bou Ammar</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> We introduce a novel approach to large language model (LLM) distillation by formulating it as a constrained reinforcement learning problem. While recent work has begun exploring the integration of task-specific rewards into distillation processes, existing methods typically rely on ad-hoc reward weighting. We propose a principled optimization framework that maximizes task-specific rewards while constraining the divergence from the teacher model to remain below a specified threshold. Our approach adapts constrained state augmented reinforcement learning to the distillation setting, introducing a modified reward function that maintains theoretical guarantees of constraint satisfaction without requiring state augmentation or teacher model access during deployment and without the computational overhead of the dual Lagrangian methods. Through extensive experiments on mathematical reasoning tasks, we demonstrate that our method achieves better constraint satisfaction rates and better reasoning compared to the soft Lagrangian relaxation baselines while maintaining competitive task performance. Our framework provides a theoretically grounded and practically efficient solution for reward-aware distillation in resource-constrained settings.</p> - https://arxiv.org/abs/2509.22921 - Fri, 26 Sep 2025 20:47:49 +0000 + Humanline: Online Alignment as Perceptual Loss + https://arxiv.org/abs/2509.24207 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24207.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sijia Liu, Niklas Muennighoff, Kawin Ethayarajh</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Online alignment (e.g., GRPO) is generally more performant than offline alignment (e.g., DPO) -- but why? Drawing on prospect theory from behavioral economics, we propose a human-centric explanation. We prove that online on-policy sampling better approximates the human-perceived distribution of what the model can produce, and PPO/GRPO-style clipping -- originally introduced to just stabilize training -- recovers a perceptual bias in how humans perceive probability. In this sense, PPO/GRPO act as perceptual losses already. Our theory further suggests that the online/offline dichotomy is itself incidental to maximizing human utility, since we can achieve the same effect by selectively training on any data in a manner that mimics human perception, rather than restricting ourselves to online on-policy data. Doing so would allow us to post-train more quickly, cheaply, and flexibly without sacrificing performance. To this end, we propose a design pattern that explicitly incorporates perceptual distortions of probability into objectives like DPO/KTO/GRPO, creating humanline variants of them. Surprisingly, we find that these humanline variants, even when trained with offline off-policy data, can match the performance of their online counterparts on both verifiable and unverifiable tasks.</p> + https://arxiv.org/abs/2509.24207 + Mon, 29 Sep 2025 02:41:16 +0000 - ADAM: A Diverse Archive of Mankind for Evaluating and Enhancing LLMs in Biographical Reasoning - https://arxiv.org/abs/2509.22991 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.22991.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jasin Cekinmez, Omid Ghahroodi, Saad Fowad Chandle, Dhiman Gupta, Ehsaneddin Asgari</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> We introduce ADAM (A Diverse Archive of Mankind), a framework for evaluating and improving multimodal large language models (MLLMs) in biographical reasoning. To the best of our knowledge, this is the first work to systematically examine LLM capabilities in biography, a critical yet underexplored dimension of factual knowledge. At its core, AdamDB is a multilingual and multimodal dataset covering over 4 million individuals across geography, time, and profession, while AdamBench provides cognitively structured evaluations based on Bloom's taxonomy, spanning six reasoning levels in both English and native languages. To address hallucinations, particularly for lesser-known individuals, we propose AdamRAG, a retrieval-augmented generation system tailored to biographical contexts. Experiments show that AdamRAG substantially improves open-source models and modestly benefits closed-source ones, with the largest gains on lower-order reasoning. Popularity strongly mediates accuracy, and multimodal input via face images offers smaller, less consistent improvements than retrieval. ADAM establishes the first benchmark and framework for cognitively, culturally, and multimodally grounded biographical evaluation, advancing the development of multilingual, accurate, and hallucination-resistant MLLMs.</p> - https://arxiv.org/abs/2509.22991 - Fri, 26 Sep 2025 23:04:28 +0000 + Specialization after Generalization: Towards Understanding Test-Time Training in Foundation Models + https://arxiv.org/abs/2509.24510 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24510.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jonas Hübotter, Patrik Wolf, Alexander Shevchenko, Dennis Jüni, Andreas Krause, Gil Kur</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Recent empirical studies have explored the idea of continuing to train a model at test-time for a given task, known as test-time training (TTT), and have found it to yield significant performance improvements. However, there is limited understanding of why and when TTT is effective. Earlier explanations mostly focused on the observation that TTT may help when applied to out-of-distribution adaptation or used with privileged data. However, the growing scale of foundation models with most test data being in-distribution questions these explanations. We instead posit that foundation models remain globally underparameterized, with TTT providing a mechanism for specialization after generalization, focusing capacity on concepts relevant to the test task. Specifically, under the linear representation hypothesis, we propose a model in which TTT achieves a substantially smaller in-distribution test error than global training. We empirically validate our model's key assumptions by training a sparse autoencoder on ImageNet, showing that semantically related data points are explained by only a few shared concepts. Finally, we perform scaling studies across image and language tasks that confirm the practical implications of our model, identifying the regimes where specialization is most effective.</p> + https://arxiv.org/abs/2509.24510 + Mon, 29 Sep 2025 09:24:52 +0000 - Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding - https://arxiv.org/abs/2509.23050 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23050.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lin Long, Changdae Oh, Seongheon Park, Yixuan Li</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Large vision-language models (LVLMs) achieve strong performance on multimodal tasks, yet they often default to their language prior (LP) -- memorized textual patterns from pre-training while under-utilizing visual evidence. Prior analyses of LP mostly rely on input-output probing, which fails to reveal the internal mechanisms governing when and how vision influences model behavior. To address this gap, we present the first systematic analysis of language prior through the lens of chain-of-embedding, which examines the layer-wise representation dynamics within LVLMs. Our analysis reveals a universal phenomenon: each model exhibits a Visual Integration Point (VIP), a critical layer at which visual information begins to meaningfully reshape hidden representations and influence decoding. Building on this observation, we introduce the Total Visual Integration (TVI) estimator, which aggregates representation distance beyond the VIP to quantify how strongly visual query influences response generation. Across 54 model-dataset combinations spanning 9 contemporary LVLMs and 6 benchmarks, we demonstrate that VIP consistently emerges, and that TVI reliably predicts the strength of language prior. This offers a principled toolkit for diagnosing and understanding language prior in LVLMs.</p> - https://arxiv.org/abs/2509.23050 - Sat, 27 Sep 2025 02:12:05 +0000 + Who invented deep residual learning? + https://arxiv.org/abs/2509.24732 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24732.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Juergen Schmidhuber</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Modern AI is based on deep artificial neural networks (NNs). As of 2025, the most cited scientific article of the 21st century is an NN paper on deep residual learning with residual connections. Who invented this? We present a timeline of the evolution of deep residual learning.</p> + https://arxiv.org/abs/2509.24732 + Mon, 29 Sep 2025 12:57:35 +0000 - Local Success Does Not Compose: Benchmarking Large Language Models for Compositional Formal Verification - https://arxiv.org/abs/2509.23061 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23061.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xu Xu, Xin Li, Xingwei Qu, Jie Fu, Binhang Yuan</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> We introduce DafnyCOMP, a benchmark for evaluating large language models (LLMs) on compositional specification generation in Dafny. Unlike prior benchmarks that focus on single-function tasks, DafnyCOMP targets programs composed of multiple interacting functions with data dependencies, requiring reasoning across component boundaries. The benchmark consists of 300 automatically synthesized multi-function programs. We evaluate several state-of-the-art LLM families and find that, while they perform well on single-function verification, their performance drops sharply on compositional tasks. Analysis reveals systematic failures in cross-functional reasoning, including fragile specifications, misalignment between implementations and proofs, and unstable reasoning. DafnyCOMP thus provides a diagnostic tool for measuring progress toward reliable, verifiable, and compositional code generation with LLMs.</p> - https://arxiv.org/abs/2509.23061 - Sat, 27 Sep 2025 02:33:08 +0000 + MANI-Pure: Magnitude-Adaptive Noise Injection for Adversarial Purification + https://arxiv.org/abs/2509.25082 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25082.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xiaoyi Huang, Junwei Wu, Kejia Zhang, Carl Yang, Zhiming Luo</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Adversarial purification with diffusion models has emerged as a promising defense strategy, but existing methods typically rely on uniform noise injection, which indiscriminately perturbs all frequencies, corrupting semantic structures and undermining robustness. Our empirical study reveals that adversarial perturbations are not uniformly distributed: they are predominantly concentrated in high-frequency regions, with heterogeneous magnitude intensity patterns that vary across frequencies and attack types. Motivated by this observation, we introduce MANI-Pure, a magnitude-adaptive purification framework that leverages the magnitude spectrum of inputs to guide the purification process. Instead of injecting homogeneous noise, MANI-Pure adaptively applies heterogeneous, frequency-targeted noise, effectively suppressing adversarial perturbations in fragile high-frequency, low-magnitude bands while preserving semantically critical low-frequency content. Extensive experiments on CIFAR-10 and ImageNet-1K validate the effectiveness of MANI-Pure. It narrows the clean accuracy gap to within 0.59 of the original classifier, while boosting robust accuracy by 2.15, and achieves the top-1 robust accuracy on the RobustBench leaderboard, surpassing the previous state-of-the-art method.</p> + https://arxiv.org/abs/2509.25082 + Mon, 29 Sep 2025 17:22:40 +0000 - Multiplayer Nash Preference Optimization - https://arxiv.org/abs/2509.23102 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23102.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Fang Wu, Xu Huang, Weihao Xuan, Zhiwei Zhang, Yijia Xiao, Guancheng Wan, Xiaomin Li, Bing Hu, Peng Xia, Jure Leskovec, Yejin Choi</p><p><b>Upvotes:</b> 52</p><p><b>Summary:</b> Reinforcement learning from human feedback (RLHF) has emerged as the standard paradigm for aligning large language models (LLMs) with human preferences. However, reward-based methods built on the Bradley-Terry assumption struggle to capture the non-transitive and heterogeneous nature of real-world preferences. To address this, recent studies have reframed alignment as a two-player Nash game, giving rise to Nash learning from human feedback (NLHF). While this perspective has inspired algorithms such as INPO, ONPO, and EGPO with strong theoretical and empirical guarantees, they remain fundamentally restricted to two-player interactions, creating a single-opponent bias that fails to capture the full complexity of realistic preference structures. In this work, we introduce Multiplayer Nash Preference Optimization (MNPO), a novel framework that generalizes NLHF to the multiplayer regime. It formulates alignment as an n-player game, where each policy competes against a population of opponents while being regularized toward a reference model. Our framework establishes well-defined Nash equilibria in multiplayer settings and extends the concept of duality gap to quantify approximation quality. We demonstrate that MNPO inherits the equilibrium guarantees of two-player methods while enabling richer competitive dynamics and improved coverage of diverse preference structures. Through comprehensive empirical evaluation, we show that MNPO consistently outperforms existing NLHF baselines on instruction-following benchmarks, achieving superior alignment quality under heterogeneous annotator conditions and mixed-policy evaluation scenarios. Together, these results establish MNPO as a principled and scalable framework for aligning LLMs with complex, non-transitive human preferences. Code is available at https://github.com/smiles724/MNPO.</p> - https://arxiv.org/abs/2509.23102 - Sat, 27 Sep 2025 04:18:33 +0000 + jina-reranker-v3: Last but Not Late Interaction for Document Reranking + https://arxiv.org/abs/2509.25085 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25085.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Feng Wang, Yuqing Li, Han Xiao</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> jina-reranker-v3 is a 0.6B parameter multilingual document reranker that introduces a novel last but not late interaction. Unlike late interaction models such as ColBERT that perform separate encoding followed by multi-vector matching, our approach conducts causal self-attention between query and documents within the same context window, enabling rich cross-document interactions before extracting contextual embeddings from the last token of each document. This compact architecture achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being ten times smaller than generative listwise rerankers.</p> + https://arxiv.org/abs/2509.25085 + Mon, 29 Sep 2025 17:23:54 +0000 - RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility - https://arxiv.org/abs/2509.23115 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23115.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haoyu He, Haozheng Luo, Yan Chen, Qi R. Wang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Predicting human mobility is inherently challenging due to complex long-range dependencies and multi-scale periodic behaviors. To address this, we introduce RHYTHM (Reasoning with Hierarchical Temporal Tokenization for Human Mobility), a unified framework that leverages large language models (LLMs) as general-purpose spatio-temporal predictors and trajectory reasoners. Methodologically, RHYTHM employs temporal tokenization to partition each trajectory into daily segments and encode them as discrete tokens with hierarchical attention that captures both daily and weekly dependencies, thereby significantly reducing the sequence length while preserving cyclical information. Additionally, we enrich token representations by adding pre-computed prompt embeddings for trajectory segments and prediction targets via a frozen LLM, and feeding these combined embeddings back into the LLM backbone to capture complex interdependencies. Computationally, RHYTHM freezes the pretrained LLM's backbone to reduce attention complexity and memory cost. We evaluate our model against state-of-the-art methods using three real-world datasets. Notably, RHYTHM achieves a 2.4% improvement in overall accuracy, a 5.0% increase on weekends, and a 24.6% reduction in training time. Code is publicly available at https://github.com/he-h/rhythm.</p> - https://arxiv.org/abs/2509.23115 - Sat, 27 Sep 2025 04:55:56 +0000 + LayerD: Decomposing Raster Graphic Designs into Layers + https://arxiv.org/abs/2509.25134 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25134.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Tomoyuki Suzuki, Kang-Jun Liu, Naoto Inoue, Kota Yamaguchi</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Designers craft and edit graphic designs in a layer representation, but layer-based editing becomes impossible once composited into a raster image. In this work, we propose LayerD, a method to decompose raster graphic designs into layers for re-editable creative workflow. LayerD addresses the decomposition task by iteratively extracting unoccluded foreground layers. We propose a simple yet effective refinement approach taking advantage of the assumption that layers often exhibit uniform appearance in graphic designs. As decomposition is ill-posed and the ground-truth layer structure may not be reliable, we develop a quality metric that addresses the difficulty. In experiments, we show that LayerD successfully achieves high-quality decomposition and outperforms baselines. We also demonstrate the use of LayerD with state-of-the-art image generators and layer-based editing.</p> + https://arxiv.org/abs/2509.25134 + Mon, 29 Sep 2025 17:50:12 +0000 - MathBode: Frequency-Domain Fingerprints of LLM Mathematical Reasoning - https://arxiv.org/abs/2509.23143 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23143.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Charles L. Wang</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> This paper presents MathBode, a dynamic diagnostic for mathematical reasoning in large language models (LLMs). Instead of one-shot accuracy, MathBode treats each parametric problem as a system: we drive a single parameter sinusoidally and fit first-harmonic responses of model outputs and exact solutions. This yields interpretable, frequency-resolved metrics -- gain (amplitude tracking) and phase (lag) -- that form Bode-style fingerprints. Across five closed-form families (linear solve, ratio/saturation, compound interest, 2x2 linear systems, similar triangles), the diagnostic surfaces systematic low-pass behavior and growing phase lag that accuracy alone obscures. We compare several models against a symbolic baseline that calibrates the instrument (G approx 1, phi approx 0). Results separate frontier from mid-tier models on dynamics, providing a compact, reproducible protocol that complements standard benchmarks with actionable measurements of reasoning fidelity and consistency. We open-source the dataset and code to enable further research and adoption.</p> - https://arxiv.org/abs/2509.23143 - Sat, 27 Sep 2025 06:06:36 +0000 + Who's Your Judge? On the Detectability of LLM-Generated Judgments + https://arxiv.org/abs/2509.25154 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25154.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dawei Li, Zhen Tan, Chengshuai Zhao, Bohan Jiang, Baixiang Huang, Pingchuan Ma, Abdullah Alnaibari, Kai Shu, Huan Liu</p><p><b>Upvotes:</b> 24</p><p><b>Summary:</b> Large Language Model (LLM)-based judgments leverage powerful LLMs to efficiently evaluate candidate content and provide judgment scores. However, the inherent biases and vulnerabilities of LLM-generated judgments raise concerns, underscoring the urgent need for distinguishing them in sensitive scenarios like academic peer reviewing. In this work, we propose and formalize the task of judgment detection and systematically investigate the detectability of LLM-generated judgments. Unlike LLM-generated text detection, judgment detection relies solely on judgment scores and candidates, reflecting real-world scenarios where textual feedback is often unavailable in the detection process. Our preliminary analysis shows that existing LLM-generated text detection methods perform poorly given their incapability to capture the interaction between judgment scores and candidate content -- an aspect crucial for effective judgment detection. Inspired by this, we introduce J-Detector, a lightweight and transparent neural detector augmented with explicitly extracted linguistic and LLM-enhanced features to link LLM judges' biases with candidates' properties for accurate detection. Experiments across diverse datasets demonstrate the effectiveness of J-Detector and show how its interpretability enables quantifying biases in LLM judges. Finally, we analyze key factors affecting the detectability of LLM-generated judgments and validate the practical utility of judgment detection in real-world scenarios.</p> + https://arxiv.org/abs/2509.25154 + Mon, 29 Sep 2025 17:54:57 +0000 - From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs - https://arxiv.org/abs/2509.23196 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23196.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haonan Wang, Weida Liang, Zihang Fu, Nie Zheng, Yifan Zhang, Yao Tong, Tongyao Zhu, Hao Jiang, Chuang Li, Jiaying Wu, Kenji Kawaguchi</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Recent reasoning LLMs (RLMs), especially those trained with verifier-based reinforcement learning, often perform worse with few-shot CoT than with direct answering. We revisit this paradox using high-quality reasoning traces from DeepSeek-R1 as demonstrations and find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal. A detailed analysis reveals two mechanisms behind this decline: (i) semantic misguidance, where high textual similarity leads the model to treat the target as the same as the exemplar and to copy intermediate steps verbatim; and (ii) strategy transfer failure, where the model struggles to extract useful reasoning strategies and apply them to target questions. Guided by these, we introduce Insight-to-Solve (I2S), a sequential test-time procedure that turns demonstrations into explicit, reusable insights and derives a target-specific reasoning trace; optionally, the reasoning is self-refined for coherence and correctness (I2S+). Extensive experiments on diverse benchmarks show that I2S and I2S+ consistently outperform both direct answering and test-time scaling baselines across open- and closed-source models. Even for GPT models, our method helps: on AIME'25, GPT-4.1 rises by +14.0%, and o1-mini improves by +2.7% on AIME and +1.7% on GPQA, indicating that in-context demonstrations can be harnessed effectively via insight-refine-solve framework.</p> - https://arxiv.org/abs/2509.23196 - Sat, 27 Sep 2025 08:59:31 +0000 + DC-VideoGen: Efficient Video Generation with Deep Compression Video Autoencoder + https://arxiv.org/abs/2509.25182 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25182.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junyu Chen, Wenkun He, Yuchao Gu, Yuyang Zhao, Jincheng Yu, Junsong Chen, Dongyun Zou, Yujun Lin, Zhekai Zhang, Muyang Li, Haocheng Xi, Ligeng Zhu, Enze Xie, Song Han, Han Cai</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> We introduce DC-VideoGen, a post-training acceleration framework for efficient video generation. DC-VideoGen can be applied to any pre-trained video diffusion model, improving efficiency by adapting it to a deep compression latent space with lightweight fine-tuning. The framework builds on two key innovations: (i) a Deep Compression Video Autoencoder with a novel chunk-causal temporal design that achieves 32x/64x spatial and 4x temporal compression while preserving reconstruction quality and generalization to longer videos; and (ii) AE-Adapt-V, a robust adaptation strategy that enables rapid and stable transfer of pre-trained models into the new latent space. Adapting the pre-trained Wan-2.1-14B model with DC-VideoGen requires only 10 GPU days on the NVIDIA H100 GPU. The accelerated models achieve up to 14.8x lower inference latency than their base counterparts without compromising quality, and further enable 2160x3840 video generation on a single GPU. Code: https://github.com/dc-ai-projects/DC-VideoGen.</p> + https://arxiv.org/abs/2509.25182 + Mon, 29 Sep 2025 17:59:31 +0000 - WirelessMathLM: Teaching Mathematical Reasoning for LLMs in Wireless Communications with Reinforcement Learning - https://arxiv.org/abs/2509.23219 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23219.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xin Li, Mengbing Liu, Yiyang Zhu, Wenhe Zhang, Li Wei, Jiancheng An, Chau Yuen</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Large language models (LLMs) excel at general mathematical reasoning but fail catastrophically on specialized technical mathematics. In wireless communications, where problems require precise manipulation of information-theoretic bounds, optimization constraints, and signal processing formulations, even state-of-the-art models struggle to achieve competent performance. We present WirelessMathLM, demonstrating that compact models (0.5B-7B parameters) can match or exceed much larger models through domain-specific reinforcement learning with verifiable rewards. Our key insight is that wireless mathematics problems possess a unique property--verifiable correctness--that enables effective reinforcement learning without human feedback. We construct WirelessMathBench-XL, a comprehensive benchmark of 4,027 problems from 970 papers. Using Group Relative Policy Optimization (GRPO) with binary verification rewards, we train models directly from base checkpoints without supervised warm-start. Our 7B model achieves 39.5% accuracy on WirelessMathBench-XL, approaching GPT-4o (40.4%) while using about 100 times fewer parameters than DeepSeek-R1 (671B, 57.4%). Remarkably, GRPO training nearly doubles performance across all model scales (0.5B +11%, 3B +103%, 7B +81%), with positive transfer to general mathematics benchmarks--our models gain +8.4 points on average across MATH, Minerva-Math, OlympiadBench, AMC, and AIME without any training on these tasks.</p> - https://arxiv.org/abs/2509.23219 - Sat, 27 Sep 2025 09:58:03 +0000 - - - Detecting Corpus-Level Knowledge Inconsistencies in Wikipedia with Large Language Models - https://arxiv.org/abs/2509.23233 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23233.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sina J. Semnani, Jirayu Burapacheep, Arpandeep Khatua, Thanawan Atchariyachanvanit, Zheng Wang, Monica S. Lam</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Wikipedia is the largest open knowledge corpus, widely used worldwide and serving as a key resource for training large language models (LLMs) and retrieval-augmented generation (RAG) systems. Ensuring its accuracy is therefore critical. But how accurate is Wikipedia, and how can we improve it? We focus on inconsistencies, a specific type of factual inaccuracy, and introduce the task of corpus-level inconsistency detection. We present CLAIRE, an agentic system that combines LLM reasoning with retrieval to surface potentially inconsistent claims along with contextual evidence for human review. In a user study with experienced Wikipedia editors, 87.5% reported higher confidence when using CLAIRE, and participants identified 64.7% more inconsistencies in the same amount of time. Combining CLAIRE with human annotation, we contribute WIKICOLLIDE, the first benchmark of real Wikipedia inconsistencies. Using random sampling with CLAIRE-assisted analysis, we find that at least 3.3% of English Wikipedia facts contradict another fact, with inconsistencies propagating into 7.3% of FEVEROUS and 4.0% of AmbigQA examples. Benchmarking strong baselines on this dataset reveals substantial headroom: the best fully automated system achieves an AUROC of only 75.1%. Our results show that contradictions are a measurable component of Wikipedia and that LLM-based systems like CLAIRE can provide a practical tool to help editors improve knowledge consistency at scale.</p> - https://arxiv.org/abs/2509.23233 - Sat, 27 Sep 2025 10:32:41 +0000 - - - Toward Effective Tool-Integrated Reasoning via Self-Evolved Preference Learning - https://arxiv.org/abs/2509.23285 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23285.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yifei Chen, Guanting Dong, Zhicheng Dou</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Tool-Integrated Reasoning (TIR) enables large language models (LLMs) to improve their internal reasoning ability by integrating external tools. However, models employing TIR often display suboptimal behaviors, such as insufficient or excessive tool usage and overthinking after tool calls. The challenge of incentivizing LLMs to perform TIR efficiently and accurately, while stabilizing the reasoning process, remains an open question. In this paper, we start by exploring the impact of tool calls on model reasoning from the perspective of information entropy. Our findings indicate that tool call results lead to a distinct change in the information entropy of subsequent reasoning, with the overall entropy of the reasoning chain varying based on the number of tool calls. Building on these insights, we propose Tool-Light, a framework designed to encourage LLMs to perform TIR efficiently and accurately. Our framework includes dataset construction and multi-stage fine-tuning. For dataset construction, we employ continuous self-evolved sampling using the fine-tuned model, integrating both vanilla sampling and entropy-guided sampling. Besides, we establish strict criteria for selecting positive-negative pairs during sampling. The training process involves a two-stage approach, comprising Supervised Fine-Tuning (SFT) and Self-Evolved Direct Preference Optimization (DPO). Experimental results on 10 datasets demonstrate the effectiveness of Tool-Light, significantly improving the model's efficiency in executing TIR tasks.</p> - https://arxiv.org/abs/2509.23285 - Sat, 27 Sep 2025 12:53:37 +0000 - - - PARROT: A Benchmark for Evaluating LLMs in Cross-System SQL Translation - https://arxiv.org/abs/2509.23338 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23338.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wei Zhou, Guoliang Li, Haoyu Wang, Yuxing Han, Xufei Wu, Fan Wu, Xuanhe Zhou</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Large language models (LLMS) have shown increasing effectiveness in Text-to-SQL tasks. However, another closely related problem, Cross-System SQL Translation (a.k.a., SQL-to-SQL), which adapts a query written for one database system (e.g., MySQL) into its equivalent one for another system (e.g., ClickHouse), is of great practical importance but remains underexplored. Existing SQL benchmarks are not well-suited for SQL-to-SQL evaluation, which (1) focus on a limited set of database systems (often just SQLite) and (2) cannot capture many system-specific SQL dialects (e.g., customized functions, data types, and syntax rules). Thus, in this paper, we introduce PARROT, a Practical And Realistic BenchmaRk for CrOss-System SQL Translation. PARROT comprises 598 translation pairs from 38 open-source benchmarks and real-world business services, specifically prepared to challenge system-specific SQL understanding (e.g., LLMS achieve lower than 38.53% accuracy on average). We also provide multiple benchmark variants, including PARROT-Diverse with 28,003 translations (for extensive syntax testing) and PARROT-Simple with 5,306 representative samples (for focused stress testing), covering 22 production-grade database systems. To promote future research, we release a public leaderboard and source code at: https://code4db.github.io/parrot-bench/.</p> - https://arxiv.org/abs/2509.23338 - Sat, 27 Sep 2025 14:41:13 +0000 - - - Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization - https://arxiv.org/abs/2509.23371 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23371.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junming Yang, Ning Xu, Biao Liu, Shiqi Qiao, Xin Geng</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Preference optimization is crucial for aligning large language models (LLMs) with human values and intentions. A significant challenge in this process is the distribution mismatch between pre-collected offline preference data and the evolving model policy. Existing methods attempt to reduce this gap using static heuristics or decoupled online sampling strategies, but they often fail to adapt to the model's dynamic learning state. To bridge this gap, we propose Meta-Weighted Adaptive Preference Optimization (MetaAPO), a novel framework that dynamically couples data generation with model training. MetaAPO employs a lightweight meta-learner, as an "alignment gap estimator", to evaluate the potential benefits of on-policy sampling in relation to offline data. This guides targeted online generation and assigns sample-wise meta-weights to the optimization objective, dynamically balancing the quality and distribution of online and offline data. Experiments on AlpacaEval 2, Arena-Hard and MT-Bench demonstrate that MetaAPO consistently outperforms existing preference optimization approaches across various settings, while reducing 42% in online annotation costs.</p> - https://arxiv.org/abs/2509.23371 - Sat, 27 Sep 2025 15:38:24 +0000 - - - Democratizing AI scientists using ToolUniverse - https://arxiv.org/abs/2509.23426 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23426.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shanghua Gao, Richard Zhu, Pengwei Sui, Zhenglun Kong, Sufian Aldogom, Yepeng Huang, Ayush Noori, Reza Shamji, Krishna Parvataneni, Theodoros Tsiligkaridis, Marinka Zitnik</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> AI scientists are emerging computational systems that serve as collaborative partners in discovery. These systems remain difficult to build because they are bespoke, tied to rigid workflows, and lack shared environments that unify tools, data, and analyses into a common ecosystem. In omics, unified ecosystems have transformed research by enabling interoperability, reuse, and community-driven development; AI scientists require comparable infrastructure. We present ToolUniverse, an ecosystem for building AI scientists from any language or reasoning model, whether open or closed. TOOLUNIVERSE standardizes how AI scientists identify and call tools, integrating more than 600 machine learning models, datasets, APIs, and scientific packages for data analysis, knowledge retrieval, and experimental design. It automatically refines tool interfaces for correct use by AI scientists, creates new tools from natural language descriptions, iteratively optimizes tool specifications, and composes tools into agentic workflows. In a case study of hypercholesterolemia, ToolUniverse was used to create an AI scientist to identify a potent analog of a drug with favorable predicted properties. The open-source ToolUniverse is available at https://aiscientist.tools.</p> - https://arxiv.org/abs/2509.23426 - Sat, 27 Sep 2025 17:38:53 +0000 - - - Clean First, Align Later: Benchmarking Preference Data Cleaning for Reliable LLM Alignment - https://arxiv.org/abs/2509.23564 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23564.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Min-Hsuan Yeh, Yixuan Li</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Human feedback plays a pivotal role in aligning large language models (LLMs) with human preferences. However, such feedback is often noisy or inconsistent, which can degrade the quality of reward models and hinder alignment. While various automated data cleaning methods have been proposed to mitigate this issue, a systematic evaluation of their effectiveness and generalizability remains lacking. To bridge this gap, we introduce the first comprehensive benchmark for evaluating 13 preference data cleaning methods in the context of LLM alignment. PrefCleanBench offers a standardized protocol to assess cleaning strategies in terms of alignment performance and generalizability across diverse datasets, model architectures, and optimization algorithms. By unifying disparate methods and rigorously comparing them, we uncover key factors that determine the success of data cleaning in alignment tasks. This benchmark lays the groundwork for principled and reproducible approaches to improving LLM alignment through better data quality-highlighting the crucial but underexplored role of data preprocessing in responsible AI development. We release modular implementations of all methods to catalyze further research: https://github.com/deeplearning-wisc/PrefCleanBench.</p> - https://arxiv.org/abs/2509.23564 - Sun, 28 Sep 2025 01:44:05 +0000 - - - Beyond the Exploration-Exploitation Trade-off: A Hidden State Approach for LLM Reasoning in RLVR - https://arxiv.org/abs/2509.23808 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23808.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Fanding Huang, Guanbo Huang, Xiao Fan, Yi He, Xiao Liang, Xiao Chen, Qinting Jiang, Faisal Nadeem Khan, Jingyan Jiang, Zhi Wang</p><p><b>Upvotes:</b> 38</p><p><b>Summary:</b> A prevailing view in Reinforcement Learning for Verifiable Rewards (RLVR) interprets recent progress through the lens of an exploration-exploitation trade-off, a perspective largely shaped by token-level metrics. We re-examine this perspective, proposing that this perceived trade-off may not be a fundamental constraint but rather an artifact of the measurement level. To investigate this, we shift the analysis to the semantically rich hidden-state space, adopting Effective Rank (ER) to quantify exploration and proposing its novel first- and second-order derivatives, named Effective Rank Velocity (ERV) and Effective Rank Acceleration (ERA), to capture exploitation dynamics. Our analysis reveals that at the hidden-state level, exploration and exploitation could be decoupled (Sec. 4). This finding reveals an opportunity to enhance both capacities simultaneously. This insight motivates our method, Velocity-Exploiting Rank-Learning (VERL), the first to operationalize the principle of synergistic exploration-exploitation enhancement by directly shaping the RL advantage function. The key innovation is leveraging the theoretically stable ERA as a predictive meta-controller to create a synergistic, dual-channel incentive structure. Instead of forcing a trade-off, VERL prospectively amplifies rewards for exploration to preempt overconfidence and reinforces exploitative gains to consolidate reasoning. Experiments across diverse LLMs and reasoning benchmarks show consistent gains, including up to 21.4% absolute accuracy improvement on the challenging Gaokao 2024 dataset.</p> - https://arxiv.org/abs/2509.23808 - Sun, 28 Sep 2025 11:14:58 +0000 - - - Efficient Multi-turn RL for GUI Agents via Decoupled Training and Adaptive Data Curation - https://arxiv.org/abs/2509.23866 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23866.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Pengxiang Li, Zechen Hu, Zirui Shang, Jingrong Wu, Yang Liu, Hui Liu, Zhi Gao, Chenrui Shi, Bofei Zhang, Zihao Zhang, Xiaochuan Shi, Zedong YU, Yuwei Wu, Xinxiao Wu, Yunde Jia, Liuyu Xiang, Zhaofeng He, Qing Li</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Vision-language model (VLM) based GUI agents show promise for automating complex desktop and mobile tasks, but face significant challenges in applying reinforcement learning (RL): (1) slow multi-turn interactions with GUI environments for policy rollout, and (2) insufficient high-quality agent-environment interactions for policy learning. To address these challenges, we propose DART, a Decoupled Agentic RL Training framework for GUI agents, which coordinates heterogeneous modules in a highly decoupled manner. DART separates the training system into four asynchronous modules: environment cluster, rollout service, data manager, and trainer. This design enables non-blocking communication, asynchronous training, rollout-wise trajectory sampling, and per-worker model synchronization, significantly improving the system efficiency: 1.6*GPU utilization for rollout, 1.9* training throughput, and 5.5* environment utilization. To facilitate effective learning from abundant samples, we introduce an adaptive data curation scheme: (1) pre-collecting successful trajectories for challenging tasks to supplement sparse success in online sampling; (2) dynamically adjusting rollout numbers and trajectory lengths based on task difficulty; (3) training selectively on high-entropy steps to prioritize critical decisions; (4) stabilizing learning via truncated importance sampling for policy mismatch between policy rollout and updating. On the OSWorld benchmark, DART-GUI-7B achieves a 42.13% task success rate, a 14.61% absolute gain over the base model, and 7.34% higher than open-source SOTA. We will fully open-source our training framework, data, and model checkpoints via computer-use-agents.github.io/dart-gui, which we believe is a timely contribution to the open-source community of agentic RL training.</p> - https://arxiv.org/abs/2509.23866 - Sun, 28 Sep 2025 13:19:20 +0000 - - - EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling - https://arxiv.org/abs/2509.23909 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23909.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xin Luo, Jiahao Wang, Chenyuan Wu, Shitao Xiao, Xiyan Jiang, Defu Lian, Jiajun Zhang, Dong Liu, Zheng liu</p><p><b>Upvotes:</b> 25</p><p><b>Summary:</b> Instruction-guided image editing has achieved remarkable progress, yet current models still face challenges with complex instructions and often require multiple samples to produce a desired result. Reinforcement Learning (RL) offers a promising solution, but its adoption in image editing has been severely hindered by the lack of a high-fidelity, efficient reward signal. In this work, we present a comprehensive methodology to overcome this barrier, centered on the development of a state-of-the-art, specialized reward model. We first introduce EditReward-Bench, a comprehensive benchmark to systematically evaluate reward models on editing quality. Building on this benchmark, we develop EditScore, a series of reward models (7B-72B) for evaluating the quality of instruction-guided image editing. Through meticulous data curation and filtering, EditScore effectively matches the performance of learning proprietary VLMs. Furthermore, coupled with an effective self-ensemble strategy tailored for the generative nature of EditScore, our largest variant even surpasses GPT-5 in the benchmark. We then demonstrate that a high-fidelity reward model is the key to unlocking online RL for image editing. Our experiments show that, while even the largest open-source VLMs fail to provide an effective learning signal, EditScore enables efficient and robust policy optimization. Applying our framework to a strong base model, OmniGen2, results in a final model that shows a substantial and consistent performance uplift. Overall, this work provides the first systematic path from benchmarking to reward modeling to RL training in image editing, showing that a high-fidelity, domain-specialized reward model is the key to unlocking the full potential of RL in this domain.</p> - https://arxiv.org/abs/2509.23909 - Sun, 28 Sep 2025 14:28:24 +0000 - - - Taming Masked Diffusion Language Models via Consistency Trajectory Reinforcement Learning with Fewer Decoding Step - https://arxiv.org/abs/2509.23924 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23924.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jingyi Yang, Guanxu Chen, Xuhao Hu, Jing Shao</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Masked diffusion language models (MDLMs) have recently emerged as a promising alternative to autoregressive (AR) language models, offering properties such as parallel decoding, flexible generation orders, and the potential for fewer inference steps. Despite these advantages, decoding strategies and reinforcement learning (RL) algorithms tailored for MDLMs remain underexplored. A naive approach is to directly transfer techniques well-established for AR models to MDLMs. However, this raises an immediate question: Is such a naive transfer truly optimal? For example, 1) Block-wise and semi-AR decoding strategies are not employed during the training of MDLMs, so why do they outperform full diffusion-style decoding during inference? 2) Applying RL algorithms designed for AR models directly to MDLMs exhibits a training-inference inconsistency, since MDLM decoding are non-causal (parallel). This results in inconsistencies between the rollout trajectory and the optimization trajectory. To address these challenges, we propose EOS Early Rejection (EOSER) and Ascending Step-Size (ASS) decoding scheduler, which unlock the potential of MDLMs to perform full diffusion-style decoding, achieving competitive performance with fewer decoding steps. Additionally, we introduce Consistency Trajectory Group Relative Policy Optimization (CJ-GRPO) for taming MDLMs, which emphasizes the consistency between rollout trajectory and optimization trajectory, and reduces the optimization errors caused by skip-step optimization. We conduct extensive experiments on reasoning tasks, such as mathematical and planning benchmarks, using LLaDA-8B-Instruct. The results demonstrate that the proposed EOSER and ASS mechanisms, together with CJ-GRPO, hold significant promise for effectively and efficiently taming MDLMs. Code: https://github.com/yjyddq/EOSER-ASS-RL.</p> - https://arxiv.org/abs/2509.23924 - Sun, 28 Sep 2025 15:01:15 +0000 - - - HunyuanImage 3.0 Technical Report - https://arxiv.org/abs/2509.23951 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.23951.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Siyu Cao, Hangting Chen, Peng Chen, Yiji Cheng, Yutao Cui, Xinchi Deng, Ying Dong, Kipper Gong, Tianpeng Gu, Xiusen Gu, Tiankai Hang, Duojun Huang, Jie Jiang, Zhengkai Jiang, Weijie Kong, Changlin Li, Donghao Li, Junzhe Li, Xin Li, Yang Li, Zhenxi Li, Zhimin Li, Jiaxin Lin, Linus, Lucaz Liu, Shu Liu, Songtao Liu, Yu Liu, Yuhong Liu, Yanxin Long, Fanbin Lu, Qinglin Lu, Yuyang Peng, Yuanbo Peng, Xiangwei Shen, Yixuan Shi, Jiale Tao, Yangyu Tao, Qi Tian, Pengfei Wan, Chunyu Wang, Kai Wang, Lei Wang, Linqing Wang, Lucas Wang, Qixun Wang, Weiyan Wang, Hao Wen, Bing Wu, Jianbing Wu, Yue Wu, Senhao Xie, Fang Yang, Miles Yang, Xiaofeng Yang, Xuan Yang, Zhantao Yang, Jingmiao Yu, Zheng Yuan, Chao Zhang, Jian-Wei Zhang, Peizhen Zhang, Shi-Xue Zhang, Tao Zhang, Weigang Zhang, Yepeng Zhang, Yingfang Zhang, Zihao Zhang, Zijian Zhang, Penghao Zhao, Zhiyuan Zhao, Xuefei Zhe, Jianchen Zhu, Zhao Zhong</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> We present HunyuanImage 3.0, a native multimodal model that unifies multimodal understanding and generation within an autoregressive framework, with its image generation module publicly available. The achievement of HunyuanImage 3.0 relies on several key components, including meticulous data curation, advanced architecture design, a native Chain-of-Thoughts schema, progressive model pre-training, aggressive model post-training, and an efficient infrastructure that enables large-scale training and inference. With these advancements, we successfully trained a Mixture-of-Experts (MoE) model comprising over 80 billion parameters in total, with 13 billion parameters activated per token during inference, making it the largest and most powerful open-source image generative model to date. We conducted extensive experiments and the results of automatic and human evaluation of text-image alignment and visual quality demonstrate that HunyuanImage 3.0 rivals previous state-of-the-art models. By releasing the code and weights of HunyuanImage 3.0, we aim to enable the community to explore new ideas with a state-of-the-art foundation model, fostering a dynamic and vibrant multimodal ecosystem. All open source assets are publicly available at https://github.com/Tencent-Hunyuan/HunyuanImage-3.0</p> - https://arxiv.org/abs/2509.23951 - Sun, 28 Sep 2025 16:14:10 +0000 - - - SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse-Linear Attention - https://arxiv.org/abs/2509.24006 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24006.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jintao Zhang, Haoxu Wang, Kai Jiang, Shuo Yang, Kaiwen Zheng, Haocheng Xi, Ziteng Wang, Hongzhou Zhu, Min Zhao, Ion Stoica, Joseph E. Gonzalez, Jun Zhu, Jianfei Chen</p><p><b>Upvotes:</b> 98</p><p><b>Summary:</b> In Diffusion Transformer (DiT) models, particularly for video generation, attention latency is a major bottleneck due to the long sequence length and the quadratic complexity. We find that attention weights can be separated into two parts: a small fraction of large weights with high rank and the remaining weights with very low rank. This naturally suggests applying sparse acceleration to the first part and low-rank acceleration to the second. Based on this finding, we propose SLA (Sparse-Linear Attention), a trainable attention method that fuses sparse and linear attention to accelerate diffusion models. SLA classifies attention weights into critical, marginal, and negligible categories, applying O(N^2) attention to critical weights, O(N) attention to marginal weights, and skipping negligible ones. SLA combines these computations into a single GPU kernel and supports both forward and backward passes. With only a few fine-tuning steps using SLA, DiT models achieve a 20x reduction in attention computation, resulting in significant acceleration without loss of generation quality. Experiments show that SLA reduces attention computation by 95% without degrading end-to-end generation quality, outperforming baseline methods. In addition, we implement an efficient GPU kernel for SLA, which yields a 13.7x speedup in attention computation and a 2.2x end-to-end speedup in video generation on Wan2.1-1.3B.</p> - https://arxiv.org/abs/2509.24006 - Sun, 28 Sep 2025 17:58:59 +0000 - - - Sequential Diffusion Language Models - https://arxiv.org/abs/2509.24007 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24007.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yangzhou Liu, Yue Cao, Hao Li, Gen Luo, Zhe Chen, Weiyun Wang, Xiaobo Liang, Biqing Qi, Lijun Wu, Changyao Tian, Yanting Zhang, Yuqiang Li, Tong Lu, Yu Qiao, Jifeng Dai, Wenhai Wang</p><p><b>Upvotes:</b> 22</p><p><b>Summary:</b> Diffusion language models (DLMs) have strong theoretical efficiency but are limited by fixed-length decoding and incompatibility with key-value (KV) caches. Block diffusion mitigates these issues, yet still enforces a fixed block size and requires expensive training. We introduce Next Sequence Prediction (NSP), which unifies next-token and next-block prediction, enabling the model to adaptively determine the generation length at each step. When the length is fixed to 1, NSP reduces to standard next-token prediction. Building on NSP, we propose Sequential Diffusion Language Model (SDLM), which can retrofit pre-trained autoregressive language models (ALMs) at minimal cost. Specifically, SDLM performs diffusion inference within fixed-size mask blocks, but dynamically decodes consecutive subsequences based on model confidence, thereby preserving KV-cache compatibility and improving robustness to varying uncertainty and semantics across the sequence. Experiments show that SDLM matches or surpasses strong autoregressive baselines using only 3.5M training samples, while achieving 2.1 higher throughput than Qwen-2.5. Notably, the SDLM-32B model delivers even more pronounced efficiency gains, demonstrating the strong scalability potential of our modeling paradigm. Project page and codes: https://github.com/OpenGVLab/SDLM</p> - https://arxiv.org/abs/2509.24007 - Sun, 28 Sep 2025 17:59:15 +0000 - - - SparseD: Sparse Attention for Diffusion Language Models - https://arxiv.org/abs/2509.24014 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24014.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zeqing Wang, Gongfan Fang, Xinyin Ma, Xingyi Yang, Xinchao Wang</p><p><b>Upvotes:</b> 24</p><p><b>Summary:</b> While diffusion language models (DLMs) offer a promising alternative to autoregressive models (ARs), existing open-source DLMs suffer from high inference latency. This bottleneck is mainly due to the attention's quadratic complexity with respect to context length in computing all query-key pairs. Intuitively, to reduce this complexity, a natural strategy is to restrict attention to sparse patterns that retain only the most relevant connections. Such approaches are well-established in ARs, where attention follows fixed and clearly defined sparse patterns. However, in DLMs, we observe distinct sparsity behaviors: (1) attention patterns vary across heads, (2) attention patterns in each head remain highly similar across denoising steps, and (3) early denoising steps are critical for generation. These findings render sparse attention methods designed for ARs largely incompatible with DLMs, as they fail to capture head-specific structures and risk degrading generation when applied in early denoising steps. To address these challenges, we propose SparseD, a novel sparse attention method for DLMs. Leveraging the observations, SparseD only requires pre-computing head-specific sparse patterns one time, and reuses them across all steps. This prevents recomputing sparse patterns at each denoising step. Meanwhile, SparseD uses full attention in the early steps, then switches to sparse attention later to maintain generation quality. Together, these establish SparseD as a practical and efficient solution for deploying DLMs in long-context applications. Experimental results demonstrate that SparseD achieves lossless acceleration, delivering up to 1.50times speedup over FlashAttention at a 64k context length with 1,024 denoising steps.</p> - https://arxiv.org/abs/2509.24014 - Sun, 28 Sep 2025 18:10:10 +0000 - - - TENET: Leveraging Tests Beyond Validation for Code Generation - https://arxiv.org/abs/2509.24148 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24148.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yiran Hu, Nan Jiang, Shanchao Liang, Yi Wu, Lin Tan</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Test-Driven Development (TDD) is a widely adopted software engineering practice that requires developers to create and execute tests alongside code implementation, ensuring that software behavior is continuously validated and refined. In the era of vibe coding, where developers increasingly delegate code writing to large language models (LLMs) by specifying high-level intentions, TDD becomes even more crucial, as test cases serve as executable specifications that explicitly define and verify intended functionality beyond what natural-language descriptions and code context can convey. While vibe coding under TDD is promising, there are three main challenges: (1) selecting a small yet effective test suite to improve the generation accuracy and control the execution workload, (2) retrieving context such as relevant code effectively, and (3) systematically using test feedback for effective code refinement. To address these challenges, we introduce TENET, an LLM agent for generating functions in complex real-world repositories under the TDD setting. TENET features three components: (1) a novel test harness mechanism that selects a concise test suite to maximize diversity of target usage scenarios; (2) a tailored agent toolset that performs efficient retrieval of relevant code with interactive debugging; and (3) a reflection-based refinement workflow that iteratively analyzes failures, replenishes context, and applies code refinement. TENET achieves 69.08% and 81.77% Pass@1 on RepoCod and RepoEval benchmarks, outperforming the best agentic baselines by 9.49 and 2.17 percentage points, respectively. In addition, this is the first study of test-driven code generation with repository-level context, examining how different aspects of test suites affect the performance of LLM agents under the TDD setting.</p> - https://arxiv.org/abs/2509.24148 - Mon, 29 Sep 2025 00:53:16 +0000 - - - AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play - https://arxiv.org/abs/2509.24193 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24193.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ran Xu, Yuchen Zhuang, Zihan Dong, Jonathan Wang, Yue Yu, Joyce C. Ho, Linjun Zhang, Haoyu Wang, Wenqi Shi, Carl Yang</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Search-augmented LLMs often struggle with complex reasoning tasks due to ineffective multi-hop retrieval and limited reasoning ability. We propose AceSearcher, a cooperative self-play framework that trains a single large language model (LLM) to alternate between two roles: a decomposer that breaks down complex queries and a solver that integrates retrieved contexts for answer generation. AceSearcher couples supervised fine-tuning on a diverse mixture of search, reasoning, and decomposition tasks with reinforcement fine-tuning optimized for final answer accuracy, eliminating the need for intermediate annotations. Extensive experiments on three reasoning-intensive tasks across 10 datasets show that AceSearcher outperforms state-of-the-art baselines, achieving an average exact match improvement of 7.6%. Remarkably, on document-level finance reasoning tasks, AceSearcher-32B matches the performance of the DeepSeek-V3 model using less than 5% of its parameters. Even at smaller scales (1.5B and 8B), AceSearcher often surpasses existing search-augmented LLMs with up to 9x more parameters, highlighting its exceptional efficiency and effectiveness in tackling complex reasoning tasks. Our code will be published at https://github.com/ritaranx/AceSearcher and https://huggingface.co/AceSearcher.</p> - https://arxiv.org/abs/2509.24193 - Mon, 29 Sep 2025 02:14:30 +0000 - - - UniVid: The Open-Source Unified Video Model - https://arxiv.org/abs/2509.24200 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24200.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiabin Luo, Junhui Lin, Zeyu Zhang, Biao Wu, Meng Fang, Ling Chen, Hao Tang</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Unified video modeling that combines generation and understanding capabilities is increasingly important but faces two key challenges: maintaining semantic faithfulness during flow-based generation due to text-visual token imbalance and the limitations of uniform cross-modal attention across the flow trajectory, and efficiently extending image-centric MLLMs to video without costly retraining. We present UniVid, a unified architecture that couples an MLLM with a diffusion decoder through a lightweight adapter, enabling both video understanding and generation. We introduce Temperature Modality Alignment to improve prompt adherence and Pyramid Reflection for efficient temporal reasoning via dynamic keyframe selection. Extensive experiments on standard benchmarks demonstrate state-of-the-art performance, achieving a 2.2% improvement on VBench-Long total score compared to EasyAnimateV5.1, and 1.0% and 3.3% accuracy gains on MSVD-QA and ActivityNet-QA, respectively, compared with the best prior 7B baselines.</p> - https://arxiv.org/abs/2509.24200 - Mon, 29 Sep 2025 02:31:36 +0000 - - - AdvChain: Adversarial Chain-of-Thought Tuning for Robust Safety Alignment of Large Reasoning Models - https://arxiv.org/abs/2509.24269 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24269.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zihao Zhu, Xinyu Wu, Gehan Hu, Siwei Lyu, Ke Xu, Baoyuan Wu</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in complex problem-solving through Chain-of-Thought (CoT) reasoning. However, the multi-step nature of CoT introduces new safety challenges that extend beyond conventional language model alignment. We identify a failure mode in current safety CoT tuning methods: the snowball effect, where minor reasoning deviations progressively amplify throughout the thought process, leading to either harmful compliance or excessive refusal. This effect stems from models being trained to imitate perfect reasoning scripts without learning to self-correct. To address this limitation, we propose AdvChain, an alignment paradigm that teaches models dynamic self-correction through adversarial CoT tuning. Our method involves constructing a dataset containing Temptation-Correction and Hesitation-Correction samples, where models learn to recover from harmful reasoning drifts and unnecessary cautions. Extensive experiments show that AdvChain significantly enhances robustness against jailbreak attacks and CoT hijacking while substantially reducing over-refusal on benign prompts, achieving a superior safety-utility balance without compromising reasoning capabilities. Our work establishes a new direction for building more robust and reliable reasoning models.</p> - https://arxiv.org/abs/2509.24269 - Mon, 29 Sep 2025 04:27:23 +0000 - - - SCI-Verifier: Scientific Verifier with Thinking - https://arxiv.org/abs/2509.24285 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24285.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shenghe Zheng, Chenyu Huang, Fangchen Yu, Junchi Yao, Jingqi Ye, Tao Chen, Yun Luo, Ning Ding, LEI BAI, Ganqu Cui, Peng Ye</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> As large language models (LLMs) are increasingly applied to scientific reasoning, the complexity of answer formats and the diversity of equivalent expressions make answer verification a critical yet challenging task. Existing verification studies in scientific domains suffer from two major limitations: (a) the absence of systematic evaluation standards and insufficient disciplinary coverage, which hinders their comprehensive assessment; and (b) heavy reliance on cumbersome rule design or prompt engineering, which reduces their effectiveness in complex reasoning scenarios or limits their cross-disciplinary generalization. To address these challenges, we propose solutions at both the data and model levels. On the data side, we construct SCI-VerifyBench, a cross-disciplinary benchmark covering mathematics, physics, biology, chemistry, and general scientific QA. The benchmark is built from real LLM responses and enhanced with domain-specific equivalence transformations that generate challenging and realistic data. Model-based and expert annotations ensure both quality and diversity, enabling rigorous evaluation of verification ability. On the model side, we emphasize the importance of reasoning for verification and introduce SCI-Verifier, a unified reasoning-augmented verifier for scientific domains. Through post-training, SCI-Verifier demonstrates strong logical reasoning and equivalence judgment capabilities while maintaining concise and stable outputs. Together, SCI-VerifyBench and SCI-Verifier provide a principled framework for scientific verification, offering both systematic evaluation and practical pathways to enhance the reliability and applicability of LLMs in scientific domains.</p> - https://arxiv.org/abs/2509.24285 - Mon, 29 Sep 2025 04:58:43 +0000 - - - Rethinking JEPA: Compute-Efficient Video SSL with Frozen Teachers - https://arxiv.org/abs/2509.24317 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24317.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xianhang Li, Chen Huang, Chun-Liang Li, Eran Malach, Josh Susskind, Vimal Thilak, Etai Littwin</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Video Joint Embedding Predictive Architectures (V-JEPA) learn generalizable off-the-shelf video representation by predicting masked regions in latent space with an exponential moving average (EMA)-updated teacher. While EMA prevents representation collapse, it complicates scalable model selection and couples teacher and student architectures. We revisit masked-latent prediction and show that a frozen teacher suffices. Concretely, we (i) train a target encoder with a simple pixel-reconstruction objective under V-JEPA masking, then (ii) freeze it and train a student to predict the teacher's latents on masked regions. This leads to a two-stage, unregularized scheme that we refer to as SALT (Static-teacher Asymmetric Latent Training). SALT decouples optimization into pixel reconstruction (teacher) and masked latent prediction (student), increasing transparency, efficiency, and scalability while preserving the ability of representation to generalize under frozen evaluation. Empirically, our student models outperform recently proposed V-JEPA 2 encoders under frozen backbone evaluation across diverse benchmarks. They are also more compute-optimal: at matched pretraining FLOPs, our method achieves higher probing accuracy, and its scaling curves dominate V-JEPA's accuracy-FLOPs Pareto frontier. Finally, we find that student quality is remarkably robust to teacher quality: high-performing students emerge even with small, sub-optimal teachers. This points to a compute budget allocation that should overwhelmingly favor the student. These results position SALT as a simple, scalable, and compute-efficient alternative to EMA-based self-distillation for video representation learning.</p> - https://arxiv.org/abs/2509.24317 - Mon, 29 Sep 2025 05:55:17 +0000 - - - Hyperspherical Latents Improve Continuous-Token Autoregressive Generation - https://arxiv.org/abs/2509.24335 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24335.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Guolin Ke, Hui Xue</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Autoregressive (AR) models are promising for image generation, yet continuous-token AR variants often trail latent diffusion and masked-generation models. The core issue is heterogeneous variance in VAE latents, which is amplified during AR decoding, especially under classifier-free guidance (CFG), and can cause variance collapse. We propose SphereAR to address this issue. Its core design is to constrain all AR inputs and outputs -- including after CFG -- to lie on a fixed-radius hypersphere (constant ell_2 norm), leveraging hyperspherical VAEs. Our theoretical analysis shows that hyperspherical constraint removes the scale component (the primary cause of variance collapse), thereby stabilizing AR decoding. Empirically, on ImageNet generation, SphereAR-H (943M) sets a new state of the art for AR models, achieving FID 1.34. Even at smaller scales, SphereAR-L (479M) reaches FID 1.54 and SphereAR-B (208M) reaches 1.92, matching or surpassing much larger baselines such as MAR-H (943M, 1.55) and VAR-d30 (2B, 1.92). To our knowledge, this is the first time a pure next-token AR image generator with raster order surpasses diffusion and masked-generation models at comparable parameter scales.</p> - https://arxiv.org/abs/2509.24335 - Mon, 29 Sep 2025 06:34:24 +0000 - - - Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning - https://arxiv.org/abs/2509.24372 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24372.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xin Qiu, Yulu Gan, Conor F. Hayes, Qiyao Liang, Elliot Meyerson, Babak Hodjat, Risto Miikkulainen</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Fine-tuning pre-trained large language models (LLMs) for down-stream tasks is a critical step in the AI deployment pipeline. Reinforcement learning (RL) is arguably the most prominent fine-tuning method, contributing to the birth of many state-of-the-art LLMs. In contrast, evolution strategies (ES), which once showed comparable performance to RL on models with a few million parameters, was neglected due to the pessimistic perception of its scalability to larger models. In this work, we report the first successful attempt to scale up ES for fine-tuning the full parameters of LLMs, showing the surprising fact that ES can search efficiently over billions of parameters and outperform existing RL fine-tuning methods in multiple respects, including sample efficiency, tolerance to long-horizon rewards, robustness to different base LLMs, less tendency to reward hacking, and more stable performance across runs. It therefore serves as a basis to unlock a new direction in LLM fine-tuning beyond what current RL techniques provide. The source codes are provided at: https://github.com/VsonicV/es-fine-tuning-paper.</p> - https://arxiv.org/abs/2509.24372 - Mon, 29 Sep 2025 07:19:34 +0000 - - - Euclid's Gift: Enhancing Spatial Perception and Reasoning in Vision-Language Models via Geometric Surrogate Tasks - https://arxiv.org/abs/2509.24473 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24473.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shijie Lian, Changti Wu, Laurence Tianruo Yang, Hang Yuan, Bin Yu, Lei Zhang, Kai Chen</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> Spatial intelligence spans a rich suite of abilities, including visualising and transforming shapes, mentally rotating objects, judging relational positions and containment, and estimating numerosity. However, it still remains a critical unresolved challenge for Multimodal Large Language Models (MLLMs).To fill this gap, we propose to treat Euclidean geometry problem-solving as a surrogate task. Specifically, we meticulously constructed a curated multimodal dataset, called Euclid30K, comprising approximately 30K plane and solid geometry problems. To enable the model to acquire and apply Euclidean principles from these geometry problems, we employed Group Relative Policy Optimization (GRPO) to finetune the Qwen2.5VL family and RoboBrain2.0 family, inspiring the models to identify shapes, count, and relate entities, and perform multi-step deductive reasoning using Euclidean principles. Our experiments demonstrate that the resulting models achieve substantial zero-shot gains across four spatial reasoning benchmarks (Super-CLEVR, Omni3DBench, VSI-Bench, and MindCube) without any task-specific adaptations. Notably, after training on the Euclid30K, the mean VSI-Bench accuracy of all evaluated models rose from 34.5% to 40.5%, improving by 5.5 percentage points. Among them, RoboBrain2.0-Euclid-7B achieves 49.6\% accuracy, surpassing the previous state-of-the-art model, Spatial-MLLM.To our knowledge, this is the first systematic study showing that geometry-centric fine-tuning can confer vision-language models with broadly transferable spatial skills. Code and Euclid30K dataset can be found in https://zgca-ai4edu.github.io/Euclids_Gift.</p> - https://arxiv.org/abs/2509.24473 - Mon, 29 Sep 2025 08:49:21 +0000 - - - GRPO-MA: Multi-Answer Generation in GRPO for Stable and Efficient Chain-of-Thought Training - https://arxiv.org/abs/2509.24494 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24494.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hongcheng Wang, Yinuo Huang, Sukai Wang, Guanghui Ren, Hao Dong</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Recent progress, such as DeepSeek-R1, has shown that the GRPO algorithm, a Reinforcement Learning (RL) approach, can effectively train Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs) and Vision-Language Models (VLMs). In this paper, we analyze three challenges of GRPO: gradient coupling between thoughts and answers, sparse reward signals caused by limited parallel sampling, and unstable advantage estimation. To mitigate these challenges, we propose GRPO-MA, a simple yet theoretically grounded method that leverages multi-answer generation from each thought process, enabling more robust and efficient optimization. Theoretically, we show that the variance of thought advantage decreases as the number of answers per thought increases. Empirically, our gradient analysis confirms this effect, showing that GRPO-MA reduces gradient spikes compared to GRPO. Experiments on math, code, and diverse multimodal tasks demonstrate that GRPO-MA substantially improves performance and training efficiency. Our ablation studies further reveal that increasing the number of answers per thought consistently enhances model performance.</p> - https://arxiv.org/abs/2509.24494 - Mon, 29 Sep 2025 09:07:45 +0000 - - - BPMN Assistant: An LLM-Based Approach to Business Process Modeling - https://arxiv.org/abs/2509.24592 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24592.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Josip Tomo Licardo, Nikola Tankovic, Darko Etinger</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> This paper presents BPMN Assistant, a tool that leverages Large Language Models (LLMs) for natural language-based creation and editing of BPMN diagrams. A specialized JSON-based representation is introduced as a structured alternative to the direct handling of XML to enhance the accuracy of process modifications. Process generation quality is evaluated using Graph Edit Distance (GED) and Relative Graph Edit Distance (RGED), while editing performance is evaluated with a binary success metric. Results show that JSON and XML achieve similar similarity scores in generation, but JSON offers greater reliability, faster processing, and significantly higher editing success rates. We discuss key trade-offs, limitations, and future improvements. The implementation is available at https://github.com/jtlicardo/bpmn-assistant.</p> - https://arxiv.org/abs/2509.24592 - Mon, 29 Sep 2025 10:56:08 +0000 - - - InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation - https://arxiv.org/abs/2509.24663 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24663.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Weilin Zhao, Zihan Zhou, Zhou Su, Chaojun Xiao, Yuxuan Li, Yanghao Li, Yudi Zhang, Weilun Zhao, Zhen Li, Yuxiang Huang, Ao Sun, Xu Han, Zhiyuan Liu</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Long-sequence processing is a critical capability for modern large language models. However, the self-attention mechanism in the standard Transformer architecture faces severe computational and memory bottlenecks when processing long sequences. While trainable sparse attention methods offer a promising solution, existing approaches such as NSA introduce excessive extra parameters and disrupt the conventional pretrain-on-short, finetune-on-long workflow, resulting in slow convergence and difficulty in acceleration. To overcome these limitations, we introduce dense-sparse switchable attention framework, termed as InfLLM-V2. InfLLM-V2 is a trainable sparse attention that seamlessly adapts models from short to long sequences. Specifically, InfLLM-V2 reuses dense attention parameters through parameter-free architecture modification, maintaining consistency between short and long sequence processing. Additionally, InfLLM-V2 ensures computational efficiency across all sequence lengths, by using dense attention for short inputs and smoothly transitioning to sparse attention for long sequences. To achieve practical acceleration, we further introduce an efficient implementation of InfLLM-V2 that significantly reduces the computational overhead. Our experiments on long-context understanding and chain-of-thought reasoning demonstrate that InfLLM-V2 is 4times faster than dense attention while retaining 98.1% and 99.7% of the performance, respectively. Based on the InfLLM-V2 framework, we have trained and open-sourced MiniCPM4.1 (https://huggingface.co/openbmb/MiniCPM4.1-8B), a hybrid reasoning model, providing a reproducible implementation for the research community.</p> - https://arxiv.org/abs/2509.24663 - Mon, 29 Sep 2025 12:08:33 +0000 - - - SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer - https://arxiv.org/abs/2509.24695 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24695.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie</p><p><b>Upvotes:</b> 35</p><p><b>Summary:</b> We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation.</p> - https://arxiv.org/abs/2509.24695 - Mon, 29 Sep 2025 12:28:09 +0000 - - - IWR-Bench: Can LVLMs reconstruct interactive webpage from a user interaction video? - https://arxiv.org/abs/2509.24709 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24709.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yang Chen, Minghao Liu, Yufan Shen, Yunwen Li, Tianyuan Huang, Xinyu Fang, Tianyu Zheng, Wenxuan Huang, Cheng Yang, Daocheng Fu, Jianbiao Mei, Rong Wu, Licheng Wen, Xuemeng Yang, Song Mao, Qunshu Lin, Zhi Yu, Yongliang Shen, Yu Qiao, Botian Shi</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> The webpage-to-code task requires models to understand visual representations of webpages and generate corresponding code. However, existing benchmarks primarily focus on static screenshot-to-code tasks, thereby overlooking the dynamic interactions fundamental to real-world web applications. To address this limitation, this paper introduces IWR-Bench, a novel benchmark for evaluating the capabilities of Large Vision-Language Models (LVLMs) in interactive webpage reconstruction from video. IWR-Bench comprises 113 meticulously curated tasks from 100 real-world websites, with 1,001 actions and featuring diverse interaction complexities (e.g., web games), visual styles, and domains. Aligning with standard web development practices, each task includes not only user interaction videos but also all crawled static assets (e.g., images, videos). This benchmark evaluates models on two fundamental challenges: comprehensive multi-modal reasoning to infer interaction logic from video and assets, and advanced code generation to translate this logic into functional code. An agent-as-a-judge framework with a comprehensive metric system automatically assesses the functional correctness and visual fidelity of generated webpages. Extensive experiments on 28 LVLMs reveal a significant challenge: the best model achieves an overall score of only 36.35%, as functional correctness (24.39% IFS) lags significantly behind visual fidelity (64.25% VFS). These results highlight critical limitations in current models' ability to reason about temporal dynamics and synthesize event-driven logic, establishing IWR-Bench as a challenging frontier for vision-language research. The benchmark and evaluation code will be made publicly available. Code is available at https://github.com/L-O-I/IWR-Bench.</p> - https://arxiv.org/abs/2509.24709 - Mon, 29 Sep 2025 12:38:06 +0000 - - - Socratic-Zero : Bootstrapping Reasoning via Data-Free Agent Co-evolution - https://arxiv.org/abs/2509.24726 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24726.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shaobo Wang, Zhengbo Jiao, Zifan Zhang, Yilang Peng, Xu Ze, Boyu Yang, Wei Wang, Hu Wei, Linfeng Zhang</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Recent breakthroughs in large language models (LLMs) on reasoning tasks rely heavily on massive, high-quality datasets-typically human-annotated and thus difficult to scale. While data synthesis or distillation offers a promising alternative, existing methods struggle with inconsistent data quality and an inability to dynamically adapt to the evolving capabilities of the model, leading to suboptimal training signals. To address these limitations, we introduce Socratic-Zero, a fully autonomous framework that generates high-quality training data from minimal seed examples through the co-evolution of three agents: the Teacher, the Solver, and the Generator. The Solver continuously refines its reasoning by learning from preference feedback on both successful and failed trajectories; the Teacher adaptively crafts increasingly challenging questions based on the Solver's weaknesses; and the Generator distills the Teacher's question-design strategy to enable scalable, high-fidelity curriculum generation. This closed-loop system produces a self-improving curriculum-requiring no pre-existing tasks or labels. Remarkably, starting from only 100 seed questions, our Socratic-Solver-8B achieves an average gain of +20.2 percentage points over prior data synthesis methods across seven mathematical reasoning benchmarks (AMC23, AIME24-25, Olympiad, MATH-500, Minerva, and GSM8K), with consistent gains on both Qwen3 and GLM4 series models. Even more surprisingly, synthetic data from Socratic-Generator-32B enables student LLMs to achieve superior performance compared to other state-of-the-art (SOTA) commercial LLMs on these benchmarks, including Qwen3-235B-A22B, DeepSeek-V3.1-671B, GPT-5, Gemini-2.5-Pro, Grok-4, and Claude-4.1-Opus.</p> - https://arxiv.org/abs/2509.24726 - Mon, 29 Sep 2025 12:54:07 +0000 - - - LOVE-R1: Advancing Long Video Understanding with an Adaptive Zoom-in Mechanism via Multi-Step Reasoning - https://arxiv.org/abs/2509.24786 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24786.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shenghao Fu, Qize Yang, Yuan-Ming Li, Xihan Wei, Xiaohua Xie, Wei-Shi Zheng</p><p><b>Upvotes:</b> 5</p><p><b>Summary:</b> Long video understanding is still challenging for recent Large Video-Language Models (LVLMs) due to the conflict between long-form temporal understanding and detailed spatial perception. LVLMs with a uniform frame sampling mechanism, which samples frames with an equal frame size and fixed sampling rate, inevitably sacrifice either temporal clues or spatial details, resulting in suboptimal solutions. To mitigate this dilemma, we propose LOVE-R1, a model that can adaptively zoom in on a video clip. The model is first provided with densely sampled frames but in a small resolution. If some spatial details are needed, the model can zoom in on a clip of interest with a large frame resolution based on its reasoning until key visual information is obtained. The whole process is implemented as a multi-step reasoning process. To train the reasoning ability, we first finetune the model on our collected 38k high-quality CoT data and enhance it with decoupled reinforcement finetuning. As outcome rewards can not provide fine-grained process supervision, we decouple multi-step reasoning into multiple single-step reasoning and optimize the internal zoom-in ability explicitly. Experiments on long video understanding benchmarks show that our model with the slow-fast adaptive frame sampling mechanism achieves a great trade-off between sampling density and frame resolutions, and LOVE-R1 outperforms our baseline Qwen2.5-VL by an average of 3.1% points across 4 common long video understanding benchmarks.</p> - https://arxiv.org/abs/2509.24786 - Mon, 29 Sep 2025 13:43:55 +0000 - - - ThermalGen: Style-Disentangled Flow-Based Generative Models for RGB-to-Thermal Image Translation - https://arxiv.org/abs/2509.24878 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24878.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiuhong Xiao, Roshan Nayak, Ning Zhang, Daniel Tortei, Giuseppe Loianno</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Paired RGB-thermal data is crucial for visual-thermal sensor fusion and cross-modality tasks, including important applications such as multi-modal image alignment and retrieval. However, the scarcity of synchronized and calibrated RGB-thermal image pairs presents a major obstacle to progress in these areas. To overcome this challenge, RGB-to-Thermal (RGB-T) image translation has emerged as a promising solution, enabling the synthesis of thermal images from abundant RGB datasets for training purposes. In this study, we propose ThermalGen, an adaptive flow-based generative model for RGB-T image translation, incorporating an RGB image conditioning architecture and a style-disentangled mechanism. To support large-scale training, we curated eight public satellite-aerial, aerial, and ground RGB-T paired datasets, and introduced three new large-scale satellite-aerial RGB-T datasets--DJI-day, Bosonplus-day, and Bosonplus-night--captured across diverse times, sensor types, and geographic regions. Extensive evaluations across multiple RGB-T benchmarks demonstrate that ThermalGen achieves comparable or superior translation performance compared to existing GAN-based and diffusion-based methods. To our knowledge, ThermalGen is the first RGB-T image translation model capable of synthesizing thermal images that reflect significant variations in viewpoints, sensor characteristics, and environmental conditions. Project page: http://xjh19971.github.io/ThermalGen</p> - https://arxiv.org/abs/2509.24878 - Mon, 29 Sep 2025 14:55:51 +0000 - - - RealUnify: Do Unified Models Truly Benefit from Unification? A Comprehensive Benchmark - https://arxiv.org/abs/2509.24897 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24897.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yang Shi, Yuhao Dong, Yue Ding, Yuran Wang, Xuanyu Zhu, Sheng Zhou, Wenting Liu, Haochen Tian, Rundong Wang, Huanqian Wang, Zuyan Liu, Bohan Zeng, Ruizhe Chen, Qixun Wang, Zhuoran Zhang, Xinlong Chen, Chengzhuo Tong, Bozhou Li, Chaoyou Fu, Qiang Liu, Haotian Wang, Wenjing Yang, Yuanxing Zhang, Pengfei Wan, Yi-Fan Zhang, Ziwei Liu</p><p><b>Upvotes:</b> 41</p><p><b>Summary:</b> The integration of visual understanding and generation into unified multimodal models represents a significant stride toward general-purpose AI. However, a fundamental question remains unanswered by existing benchmarks: does this architectural unification actually enable synergetic interaction between the constituent capabilities? Existing evaluation paradigms, which primarily assess understanding and generation in isolation, are insufficient for determining whether a unified model can leverage its understanding to enhance its generation, or use generative simulation to facilitate deeper comprehension. To address this critical gap, we introduce RealUnify, a benchmark specifically designed to evaluate bidirectional capability synergy. RealUnify comprises 1,000 meticulously human-annotated instances spanning 10 categories and 32 subtasks. It is structured around two core axes: 1) Understanding Enhances Generation, which requires reasoning (e.g., commonsense, logic) to guide image generation, and 2) Generation Enhances Understanding, which necessitates mental simulation or reconstruction (e.g., of transformed or disordered visual inputs) to solve reasoning tasks. A key contribution is our dual-evaluation protocol, which combines direct end-to-end assessment with a diagnostic stepwise evaluation that decomposes tasks into distinct understanding and generation phases. This protocol allows us to precisely discern whether performance bottlenecks stem from deficiencies in core abilities or from a failure to integrate them. Through large-scale evaluations of 12 leading unified models and 6 specialized baselines, we find that current unified models still struggle to achieve effective synergy, indicating that architectural unification alone is insufficient. These results highlight the need for new training strategies and inductive biases to fully unlock the potential of unified modeling.</p> - https://arxiv.org/abs/2509.24897 - Mon, 29 Sep 2025 15:07:28 +0000 - - - OpenGPT-4o-Image: A Comprehensive Dataset for Advanced Image Generation and Editing - https://arxiv.org/abs/2509.24900 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24900.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhihong Chen, Xuehai Bai, Yang Shi, Chaoyou Fu, Huanyu Zhang, Haotian Wang, Xiaoyan Sun, Zhang Zhang, Liang Wang, Yuanxing Zhang, Pengfei Wan, Yi-Fan Zhang</p><p><b>Upvotes:</b> 38</p><p><b>Summary:</b> The performance of unified multimodal models for image generation and editing is fundamentally constrained by the quality and comprehensiveness of their training data. While existing datasets have covered basic tasks like style transfer and simple object manipulation, they often lack the systematic structure and challenging scenarios required for real-world applications. To address this bottleneck, we introduce OpenGPT-4o-Image, a large-scale dataset constructed using a novel methodology that combines hierarchical task taxonomy with automated data generation. Our taxonomy not only includes fundamental capabilities such as text rendering and style control but also introduces highly practical yet challenging categories like scientific imagery for chemistry illustrations and complex instruction editing requiring simultaneous execution of multiple operations. Through an automated pipeline leveraging structured resource pools and GPT-4o, we generate 80k high-quality instruction-image pairs with controlled diversity, covering 11 major domains and 51 subtasks. Extensive experiments show that fine-tuning leading models on our dataset achieves significant performance gains across multiple benchmarks, with improvements of up to 18\% on editing tasks (UniWorld-V1 on ImgEdit-Bench) and 13% on generation tasks (Harmon on GenEval). Our work demonstrates that systematic data construction is key to advancing multimodal AI capabilities.</p> - https://arxiv.org/abs/2509.24900 - Mon, 29 Sep 2025 15:11:09 +0000 - - - BOE-XSUM: Extreme Summarization in Clear Language of Spanish Legal Decrees and Notifications - https://arxiv.org/abs/2509.24908 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24908.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Andrés Fernández García, Javier de la Rosa, Julio Gonzalo, Roser Morante, Enrique Amigó, Alejandro Benito-Santos, Jorge Carrillo-de-Albornoz, Víctor Fresno, Adrian Ghajari, Guillermo Marco, Laura Plaza, Eva Sánchez Salido</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> The ability to summarize long documents succinctly is increasingly important in daily life due to information overload, yet there is a notable lack of such summaries for Spanish documents in general, and in the legal domain in particular. In this work, we present BOE-XSUM, a curated dataset comprising 3,648 concise, plain-language summaries of documents sourced from Spain's ``Bolet\'{\i}n Oficial del Estado'' (BOE), the State Official Gazette. Each entry in the dataset includes a short summary, the original text, and its document type label. We evaluate the performance of medium-sized large language models (LLMs) fine-tuned on BOE-XSUM, comparing them to general-purpose generative models in a zero-shot setting. Results show that fine-tuned models significantly outperform their non-specialized counterparts. Notably, the best-performing model -- BERTIN GPT-J 6B (32-bit precision) -- achieves a 24\% performance gain over the top zero-shot model, DeepSeek-R1 (accuracies of 41.6\% vs.\ 33.5\%).</p> - https://arxiv.org/abs/2509.24908 - Mon, 29 Sep 2025 15:15:17 +0000 - - - Learning Goal-Oriented Language-Guided Navigation with Self-Improving Demonstrations at Scale - https://arxiv.org/abs/2509.24910 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24910.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Songze Li, Zun Wang, Gengze Zhou, Jialu Li, Xiangyu Zeng, Limin Wang, Yu Qiao, Qi Wu, Mohit Bansal, Yi Wang</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Goal-oriented language-guided navigation requires robust exploration capabilities for agents to navigate to specified goals in unknown environments without step-by-step instructions. Existing methods tend to exclusively utilize shortest-path trajectories, lacking effective exploration priors for training navigation agents. To address the above challenges, we present SID, a goal-oriented language-guided navigation learning approach with Self-Improving Demonstrations. Specifically, SID learns an initial agent on the shortest-path data sampled from environments and then leverages this agent to generate novel exploration trajectories. The novel rollouts provide demonstrations with stronger exploration strategies to train a better agent, which in turn produces higher-quality agent demonstrations for the next round of training. We show that this iterative self-improving pipeline readily scales to new environments, and the resulting demonstrations can be transferred across a variety of language-guided navigation tasks, elevating the performance ceiling in diverse goal-oriented navigation tasks. Extensive experiments demonstrate that SID significantly boosts the exploration capabilities and generalization of navigation agents. The resulting agent achieves new state-of-the-art performance on goal-oriented language-guided navigation tasks, including REVERIE, SOON, notably achieving a 50.9% success rate on the unseen validation splits of SOON, surpassing the prior leading approaches by a margin of 13.9%.</p> - https://arxiv.org/abs/2509.24910 - Mon, 29 Sep 2025 15:15:54 +0000 - - - Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards - https://arxiv.org/abs/2509.24981 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24981.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haoran He, Yuxiao Ye, Qingpeng Cai, Chen Hu, Binxing Jiao, Daxin Jiang, Ling Pan</p><p><b>Upvotes:</b> 21</p><p><b>Summary:</b> RL with Verifiable Rewards (RLVR) has emerged as a promising paradigm for improving the reasoning abilities of large language models (LLMs). Current methods rely primarily on policy optimization frameworks like PPO and GRPO, which follow generalized policy iteration that alternates between evaluating the current policy's value and improving the policy based on evaluation. While effective, they often suffer from training instability and diversity collapse, requiring complex heuristic tricks and careful tuning. We observe that standard RLVR in math reasoning can be formalized as a specialized finite-horizon Markov Decision Process with deterministic state transitions, tree-structured dynamics, and binary terminal rewards. Though large in scale, the underlying structure is simpler than general-purpose control settings for which popular RL algorithms (e.g., PPO) were developed, suggesting that several sophisticated techniques in existing methods may be reduced or even omitted. Based on this insight, we prove a surprising result: the optimal action can be recovered from the Q-function of a fixed uniformly random policy, thereby bypassing the generalized policy iteration loop and its associated heuristics. We introduce Random Policy Valuation for Diverse Reasoning (ROVER) to translate this principle into a practical and scalable algorithm for LLM math reasoning, a minimalist yet highly effective RL method that samples actions from a softmax over these uniform-policy Q-values. ROVER preserves diversity throughout training, allowing sustained exploration of multiple valid pathways. Across multiple base models and standard math reasoning benchmarks, ROVER demonstrates superior performance in both quality (+8.2 on pass@1, +16.8 on pass@256) and diversity (+17.6\%), despite its radical simplification compared to strong, complicated existing methods.</p> - https://arxiv.org/abs/2509.24981 - Mon, 29 Sep 2025 16:09:07 +0000 - - - Generalized Correctness Models: Learning Calibrated and Model-Agnostic Correctness Predictors from Historical Patterns - https://arxiv.org/abs/2509.24988 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.24988.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Hanqi Xiao, Vaidehi Patil, Hyunji Lee, Elias Stengel-Eskin, Mohit Bansal</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Generating accurate and calibrated confidence estimates is critical for deploying LLMs in high-stakes or user-facing applications, and remains an open challenge. Prior research has often framed confidence as a problem of eliciting a model's "self-knowledge", i.e., the ability of an LLM to judge whether its own answers are correct; this approach implicitly assumes that there is some privileged information about the answer's correctness that is accessible to the model itself. However, our experiments reveal that an LLM attempting to predict the correctness of its own outputs generally performs no better than an unrelated LLM. Moreover, we hypothesize that a key factor in building a "Correctness Model" (CM) is exposure to a target model's historical predictions. We propose multiple methods to inject this historical correctness information, creating a Generalized Correctness Model (GCM). We first show that GCMs can be trained on the correctness data from many LLMs and learn patterns for correctness prediction applicable across datasets and models. We then use CMs as a lens for studying the source of correctness prediction ability and its generalization, systematically controlling their training data and finding that answer phrasing is a strong predictor for correctness. We further explore alternative methods of injecting history without training an LLM, finding that including history as in-context examples can help improve correctness prediction, and post-hoc calibration can provide complementary reductions in calibration error. We evaluate GCMs based on Qwen3-8B across 5 model families and the MMLU and TriviaQA datasets, as well as on a downstream selective prediction task, finding that reliable LLM confidence estimation is a generalizable and model-agnostic skill learned by systematically encoding correctness history rather than a model-specific skill reliant on self-introspection.</p> - https://arxiv.org/abs/2509.24988 - Mon, 29 Sep 2025 16:19:01 +0000 - - - Advantage Weighted Matching: Aligning RL with Pretraining in Diffusion Models - https://arxiv.org/abs/2509.25050 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25050.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shuchen Xue, Chongjian Ge, Shilong Zhang, Yichen Li, Zhi-Ming Ma</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Reinforcement Learning (RL) has emerged as a central paradigm for advancing Large Language Models (LLMs), where pre-training and RL post-training share the same log-likelihood formulation. In contrast, recent RL approaches for diffusion models, most notably Denoising Diffusion Policy Optimization (DDPO), optimize an objective different from the pretraining objectives--score/flow matching loss. In this work, we establish a novel theoretical analysis: DDPO is an implicit form of score/flow matching with noisy targets, which increases variance and slows convergence. Building on this analysis, we introduce Advantage Weighted Matching (AWM), a policy-gradient method for diffusion. It uses the same score/flow-matching loss as pretraining to obtain a lower-variance objective and reweights each sample by its advantage. In effect, AWM raises the influence of high-reward samples and suppresses low-reward ones while keeping the modeling objective identical to pretraining. This unifies pretraining and RL conceptually and practically, is consistent with policy-gradient theory, reduces variance, and yields faster convergence. This simple yet effective design yields substantial benefits: on GenEval, OCR, and PickScore benchmarks, AWM delivers up to a 24times speedup over Flow-GRPO (which builds on DDPO), when applied to Stable Diffusion 3.5 Medium and FLUX, without compromising generation quality. Code is available at https://github.com/scxue/advantage_weighted_matching.</p> - https://arxiv.org/abs/2509.25050 - Mon, 29 Sep 2025 17:02:20 +0000 - - - Cogito, Ergo Ludo: An Agent that Learns to Play by Reasoning and Planning - https://arxiv.org/abs/2509.25052 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25052.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sai Wang, Yu Wu, Zhongwen Xu</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> The pursuit of artificial agents that can learn to master complex environments has led to remarkable successes, yet prevailing deep reinforcement learning methods often rely on immense experience, encoding their knowledge opaquely within neural network weights. We propose a different paradigm, one in which an agent learns to play by reasoning and planning. We introduce Cogito, ergo ludo (CEL), a novel agent architecture that leverages a Large Language Model (LLM) to build an explicit, language-based understanding of its environment's mechanics and its own strategy. Starting from a tabula rasa state with no prior knowledge (except action set), CEL operates on a cycle of interaction and reflection. After each episode, the agent analyzes its complete trajectory to perform two concurrent learning processes: Rule Induction, where it refines its explicit model of the environment's dynamics, and Strategy and Playbook Summarization, where it distills experiences into an actionable strategic playbook. We evaluate CEL on diverse grid-world tasks (i.e., Minesweeper, Frozen Lake, and Sokoban), and show that the CEL agent successfully learns to master these games by autonomously discovering their rules and developing effective policies from sparse rewards. Ablation studies confirm that the iterative process is critical for sustained learning. Our work demonstrates a path toward more general and interpretable agents that not only act effectively but also build a transparent and improving model of their world through explicit reasoning on raw experience.</p> - https://arxiv.org/abs/2509.25052 - Mon, 29 Sep 2025 17:02:31 +0000 - - - BRIDGE - Building Reinforcement-Learning Depth-to-Image Data Generation Engine for Monocular Depth Estimation - https://arxiv.org/abs/2509.25077 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25077.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Dingning Liu, Haoyu Guo, Jingyi Zhou, Tong He</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Monocular Depth Estimation (MDE) is a foundational task for computer vision. Traditional methods are limited by data scarcity and quality, hindering their robustness. To overcome this, we propose BRIDGE, an RL-optimized depth-to-image (D2I) generation framework that synthesizes over 20M realistic and geometrically accurate RGB images, each intrinsically paired with its ground truth depth, from diverse source depth maps. Then we train our depth estimation model on this dataset, employing a hybrid supervision strategy that integrates teacher pseudo-labels with ground truth depth for comprehensive and robust training. This innovative data generation and training paradigm enables BRIDGE to achieve breakthroughs in scale and domain diversity, consistently outperforming existing state-of-the-art approaches quantitatively and in complex scene detail capture, thereby fostering general and robust depth features. Code and models are available at https://dingning-liu.github.io/bridge.github.io/.</p> - https://arxiv.org/abs/2509.25077 - Mon, 29 Sep 2025 17:19:45 +0000 - - - Scaling Generalist Data-Analytic Agents - https://arxiv.org/abs/2509.25084 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25084.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shuofei Qiao, Yanqiu Zhao, Zhisong Qiu, Xiaobin Wang, Jintian Zhang, Zhao Bin, Ningyu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Data-analytic agents are emerging as a key catalyst for automated scientific discovery and for the vision of Innovating AI. Current approaches, however, rely heavily on prompt engineering over proprietary models, while open-source models struggle to face diverse-format, large-scale data files and long-horizon, multi-step reasoning that real-world analytics demands. This paper introduces DataMind, a scalable data synthesis and agent training recipe designed to build generalist data-analytic agents. DataMind tackles three key challenges in building open-source data-analytic agents, including insufficient data resources, improper training strategy, and unstable code-based multi-turn rollout. Concretely, DataMind applies 1) a fine-grained task taxonomy and a recursive easy-to-hard task composition mechanism to increase the diversity and difficulty of synthesized queries; 2) a knowledge-augmented trajectory sampling strategy followed by model-based and rule-based filtering; 3) a dynamically adjustable training objective combining both SFT and RL losses; 4) a memory-frugal and stable code-based multi-turn rollout framework. Built on DataMind, we curate DataMind-12K, a high-quality trajectory set spanning diverse domains, task categories, and data file formats for data-analytic tasks. Trained on DataMind-12K, our DataMind-14B achieves state-of-the-art with an average score of 71.16% on multiple data analysis benchmarks, outperforming the strongest proprietary baselines DeepSeek-V3.1 and GPT-5. Our DataMind-7B also performs best among all open-source models with a score of 68.10%. We also incorporate some empirical insights gained from our exploratory trials into the analysis experiments, aiming to provide actionable insights about agentic training for the community. We will release DataMind-12K and DataMind-7B,14B for the community's future research.</p> - https://arxiv.org/abs/2509.25084 - Mon, 29 Sep 2025 17:23:08 +0000 - - - Towards Personalized Deep Research: Benchmarks and Evaluations - https://arxiv.org/abs/2509.25106 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25106.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yuan Liang, Jiaxian Li, Yuqing Wang, Piaohong Wang, Motong Tian, Pai Liu, Shuofei Qiao, Runnan Fang, He Zhu, Ge Zhang, Minghao Liu, Yuchen Eleanor Jiang, Ningyu Zhang, Wangchunshu Zhou</p><p><b>Upvotes:</b> 21</p><p><b>Summary:</b> Deep Research Agents (DRAs) can autonomously conduct complex investigations and generate comprehensive reports, demonstrating strong real-world potential. However, existing evaluations mostly rely on close-ended benchmarks, while open-ended deep research benchmarks remain scarce and typically neglect personalized scenarios. To bridge this gap, we introduce Personalized Deep Research Bench, the first benchmark for evaluating personalization in DRAs. It pairs 50 diverse research tasks across 10 domains with 25 authentic user profiles that combine structured persona attributes with dynamic real-world contexts, yielding 250 realistic user-task queries. To assess system performance, we propose the PQR Evaluation Framework, which jointly measures (P) Personalization Alignment, (Q) Content Quality, and (R) Factual Reliability. Our experiments on a range of systems highlight current capabilities and limitations in handling personalized deep research. This work establishes a rigorous foundation for developing and evaluating the next generation of truly personalized AI research assistants.</p> - https://arxiv.org/abs/2509.25106 - Mon, 29 Sep 2025 17:39:17 +0000 - - - From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones - https://arxiv.org/abs/2509.25123 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25123.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Lifan Yuan, Weize Chen, Yuchen Zhang, Ganqu Cui, Hanbin Wang, Ziming You, Ning Ding, Zhiyuan Liu, Maosong Sun, Hao Peng</p><p><b>Upvotes:</b> 16</p><p><b>Summary:</b> Does RL teach LLMs genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills. To mitigate data contamination and other confounding factors, and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function f(x) given x. When an LLM has already learned f and g prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them h(x)=g(f(x)). Further, this compositional ability generalizes to more difficult problems such as compositions of >2 functions unseen during RL training. Surprisingly, our experiments show that compositional skill acquired on a source task transfers to a different target task. This transfer happens even without compositional training on the target, requiring only prior knowledge of the target's atomic skills. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, next-token training with the same data yields none of these findings. Our systematic experiments provide fresh insights into LLM learning, suggesting the value of first building base models with basic skills, then using RL to incentivize advanced, generalizable skills for complex problems.</p> - https://arxiv.org/abs/2509.25123 - Mon, 29 Sep 2025 17:44:27 +0000 - - - MGM-Omni: Scaling Omni LLMs to Personalized Long-Horizon Speech - https://arxiv.org/abs/2509.25131 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25131.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chengyao Wang, Zhisheng Zhong, Bohao Peng, Senqiao Yang, Yuqi Liu, Haokun Gui, Bin Xia, Jingyao Li, Bei Yu, Jiaya Jia</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> We present MGM-Omni, a unified Omni LLM for omni-modal understanding and expressive, long-horizon speech generation. Unlike cascaded pipelines that isolate speech synthesis, MGM-Omni adopts a "brain-mouth" design with a dual-track, token-based architecture that cleanly decouples multimodal reasoning from real-time speech generation. This design enables efficient cross-modal interaction and low-latency, streaming speech generation. For understanding, a unified training strategy coupled with a dual audio encoder design enables long-form audio perception across diverse acoustic conditions. For generation, a chunk-based parallel decoding scheme narrows the text speech token-rate gap, accelerating inference and supporting streaming zero-shot voice cloning with stable timbre over extended durations. Compared to concurrent work, MGM-Omni achieves these capabilities with markedly data-efficient training. Extensive experiments demonstrate that MGM-Omni outperforms existing open source models in preserving timbre identity across extended sequences, producing natural and context-aware speech, and achieving superior long-form audio and omnimodal understanding. MGM-Omni establishes an efficient, end-to-end paradigm for omnimodal understanding and controllable, personalised long-horizon speech generation.</p> - https://arxiv.org/abs/2509.25131 - Mon, 29 Sep 2025 17:48:28 +0000 - - - The Era of Real-World Human Interaction: RL from User Conversations - https://arxiv.org/abs/2509.25137 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25137.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chuanyang Jin, Jing Xu, Bo Liu, Leitian Tao, Olga Golovneva, Tianmin Shu, Wenting Zhao, Xian Li, Jason Weston</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> We posit that to achieve continual model improvement and multifaceted alignment, future models must learn from natural human interaction. Current conversational models are aligned using pre-annotated, expert-generated human feedback. In this work, we introduce Reinforcement Learning from Human Interaction (RLHI), a paradigm that learns directly from in-the-wild user conversations. We develop two complementary methods: (1) RLHI with User-Guided Rewrites, which revises unsatisfactory model outputs based on users' natural-language follow-up responses, (2) RLHI with User-Based Rewards, which learns via a reward model conditioned on knowledge of the user's long-term interaction history (termed persona). Together, these methods link long-term user personas to turn-level preferences via persona-conditioned preference optimization. Trained on conversations derived from WildChat, both RLHI variants outperform strong baselines in personalization and instruction-following, and similar feedback enhances performance on reasoning benchmarks. These results suggest organic human interaction offers scalable, effective supervision for personalized alignment.</p> - https://arxiv.org/abs/2509.25137 - Mon, 29 Sep 2025 17:50:31 +0000 - - - Pretraining Large Language Models with NVFP4 - https://arxiv.org/abs/2509.25149 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25149.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> NVIDIA, Felix Abecassis, Anjulie Agrusa, Dong Ahn, Jonah Alben, Stefania Alborghetti, Michael Andersch, Sivakumar Arayandi, Alexis Bjorlin, Aaron Blakeman, Evan Briones, Ian Buck, Bryan Catanzaro, Jinhang Choi, Mike Chrzanowski, Eric Chung, Victor Cui, Steve Dai, Bita Darvish Rouhani, Carlo del Mundo, Deena Donia, Burc Eryilmaz, Henry Estela, Abhinav Goel, Oleg Goncharov, Yugi Guvvala, Robert Hesse, Russell Hewett, Herbert Hum, Ujval Kapasi, Brucek Khailany, Mikail Khona, Nick Knight, Alex Kondratenko, Ronny Krashinsky, Ben Lanir, Simon Layton, Michael Lightstone, Daniel Lo, Paulius Micikevicius, Asit Mishra, Tim Moon, Deepak Narayanan, Chao Ni, Abhijit Paithankar, Satish Pasumarthi, Ankit Patel, Mostofa Patwary, Ashwin Poojary, Gargi Prasad, Sweta Priyadarshi, Yigong Qin, Xiaowei Ren, Oleg Rybakov, Charbel Sakr, Sanjeev Satheesh, Stas Sergienko, Pasha Shamis, Kirthi Shankar, Nishant Sharma, Mohammad Shoeybi, Michael Siu, Misha Smelyanskiy, Darko Stosic, Dusan Stosic, Bor-Yiing Su, Frank Sun, Nima Tajbakhsh, Shelby Thomas, Przemek Tredak, Evgeny Tsykunov, Gandhi Vaithilingam, Aditya Vavre, Rangharajan Venkatesan, Roger Waleffe, Qiyu Wan, Hexin Wang, Mengdi Wang, Lizzie Wei, Hao Wu, Evan Wu, Keith Wyss, Ning Xu, Jinze Xue, Charlene Yang, Yujia Zhai, Ruoxi Zhang, Jingyang Zhu, Zhongbo Zhu</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> Large Language Models (LLMs) today are powerful problem solvers across many domains, and they continue to get stronger as they scale in model size, training set size, and training set quality, as shown by extensive research and experimentation across the industry. Training a frontier model today requires on the order of tens to hundreds of yottaflops, which is a massive investment of time, compute, and energy. Improving pretraining efficiency is therefore essential to enable the next generation of even more capable LLMs. While 8-bit floating point (FP8) training is now widely adopted, transitioning to even narrower precision, such as 4-bit floating point (FP4), could unlock additional improvements in computational speed and resource utilization. However, quantization at this level poses challenges to training stability, convergence, and implementation, notably for large-scale models trained on long token horizons. In this study, we introduce a novel approach for stable and accurate training of large language models (LLMs) using the NVFP4 format. Our method integrates Random Hadamard transforms (RHT) to bound block-level outliers, employs a two-dimensional quantization scheme for consistent representations across both the forward and backward passes, utilizes stochastic rounding for unbiased gradient estimation, and incorporates selective high-precision layers. We validate our approach by training a 12-billion-parameter model on 10 trillion tokens -- the longest publicly documented training run in 4-bit precision to date. Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline. These findings highlight that NVFP4, when combined with our training approach, represents a major step forward in narrow-precision LLM training algorithms.</p> - https://arxiv.org/abs/2509.25149 - Mon, 29 Sep 2025 17:53:17 +0000 - - - GSM8K-V: Can Vision Language Models Solve Grade School Math Word Problems in Visual Contexts - https://arxiv.org/abs/2509.25160 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25160.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Fan Yuan, Yuchen Yan, Yifan Jiang, Haoran Zhao, Tao Feng, Jinyan Chen, Yanwei Lou, Wenqi Zhang, Yongliang Shen, Weiming Lu, Jun Xiao, Yueting Zhuang</p><p><b>Upvotes:</b> 25</p><p><b>Summary:</b> Vision language models (VLMs) achieve unified modeling of images and text, enabling them to accomplish complex real-world tasks through perception, planning, and reasoning. Among these tasks, reasoning is particularly representative, with mathematical reasoning serving as a prominent example. It highlights the high-level capability of VLMs to comprehend mathematical information in images and to perform sophisticated reasoning. Recently, numerous visual mathematical reasoning benchmarks have been proposed, but they are often restricted to geometry, lack coverage of math word problems, and rarely assess reasoning across multiple images. To address these gaps, we introduce GSM8K-V, a purely visual multi-image mathematical reasoning benchmark. GSM8K-V is built by systematically mapping each sample from the widely used text-based GSM8K into visual form. Through a carefully designed automated image-generation pipeline combined with meticulous human annotation, we curate 1,319 high-quality samples. We evaluate a wide range of open-source and closed-source models on GSM8K-V. Results show that although existing VLMs have nearly saturated performance on text-based GSM8K, there remains substantial room for improvement on GSM8K-V. For example, the best-performing model, Gemini-2.5-Pro, achieves 95.22% accuracy on GSM8K but only 46.93% on GSM8K-V. We conduct a comprehensive analysis of GSM8K-V, examining the limitations of current models as well as potential directions for improvement. GSM8K-V offers a new perspective on visual mathematical reasoning and establishes a benchmark to guide the development of more robust and generalizable VLMs.</p> - https://arxiv.org/abs/2509.25160 - Mon, 29 Sep 2025 17:57:05 +0000 - - - Rolling Forcing: Autoregressive Long Video Diffusion in Real Time - https://arxiv.org/abs/2509.25161 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25161.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Kunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan, Shijian Lu</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Streaming video generation, as one fundamental component in interactive world models and neural game engines, aims to generate high-quality, low-latency, and temporally coherent long video streams. However, most existing work suffers from severe error accumulation that often significantly degrades the generated stream videos over long horizons. We design Rolling Forcing, a novel video generation technique that enables streaming long videos with minimal error accumulation. Rolling Forcing comes with three novel designs. First, instead of iteratively sampling individual frames, which accelerates error propagation, we design a joint denoising scheme that simultaneously denoises multiple frames with progressively increasing noise levels. This design relaxes the strict causality across adjacent frames, effectively suppressing error growth. Second, we introduce the attention sink mechanism into the long-horizon stream video generation task, which allows the model to keep key value states of initial frames as a global context anchor and thereby enhances long-term global consistency. Third, we design an efficient training algorithm that enables few-step distillation over largely extended denoising windows. This algorithm operates on non-overlapping windows and mitigates exposure bias conditioned on self-generated histories. Extensive experiments show that Rolling Forcing enables real-time streaming generation of multi-minute videos on a single GPU, with substantially reduced error accumulation.</p> - https://arxiv.org/abs/2509.25161 - Mon, 29 Sep 2025 17:57:14 +0000 - - - TR2-D2: Tree Search Guided Trajectory-Aware Fine-Tuning for Discrete Diffusion - https://arxiv.org/abs/2509.25171 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25171.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sophia Tang, Yuchen Zhu, Molei Tao, Pranam Chatterjee</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Reinforcement learning with stochastic optimal control offers a promising framework for diffusion fine-tuning, where a pre-trained diffusion model is optimized to generate paths that lead to a reward-tilted distribution. While these approaches enable optimization without access to explicit samples from the optimal distribution, they require training on rollouts under the current fine-tuned model, making them susceptible to reinforcing sub-optimal trajectories that yield poor rewards. To overcome this challenge, we introduce TRee Search Guided TRajectory-Aware Fine-Tuning for Discrete Diffusion (TR2-D2), a novel framework that optimizes reward-guided discrete diffusion trajectories with tree search to construct replay buffers for trajectory-aware fine-tuning. These buffers are generated using Monte Carlo Tree Search (MCTS) and subsequently used to fine-tune a pre-trained discrete diffusion model under a stochastic optimal control objective. We validate our framework on single- and multi-objective fine-tuning of biological sequence diffusion models, highlighting the overall effectiveness of TR2-D2 for reliable reward-guided fine-tuning in discrete sequence generation.</p> - https://arxiv.org/abs/2509.25171 - Mon, 29 Sep 2025 17:58:45 +0000 - - - EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering - https://arxiv.org/abs/2509.25175 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25175.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haolei Xu, Xinyu Mei, Yuchen Yan, Rui Zhou, Wenqi Zhang, Weiming Lu, Yueting Zhuang, Yongliang Shen</p><p><b>Upvotes:</b> 23</p><p><b>Summary:</b> Large language model (LLM) steering has emerged as a promising paradigm for controlling model behavior at inference time through targeted manipulation of hidden states, offering a lightweight alternative to expensive retraining. However, existing steering frameworks suffer from critical limitations: computational inefficiency, limited extensibility, and restricted functionality that hinder both research progress and practical deployment. We present EasySteer, a unified framework for high-performance, extensible LLM steering built on vLLM. Our system features modular architecture with pluggable interfaces for both analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system. Through deep integration with vLLM's optimized inference engine, EasySteer achieves 5.5-11.4times speedup over existing frameworks. Extensive experiments demonstrate its effectiveness in overthinking mitigation, hallucination reduction, and other key applications. EasySteer transforms steering from research technique to production-ready capability, establishing critical infrastructure for deployable, controllable language models.</p> - https://arxiv.org/abs/2509.25175 - Mon, 29 Sep 2025 17:59:07 +0000 - - - SIRI: Scaling Iterative Reinforcement Learning with Interleaved Compression - https://arxiv.org/abs/2509.25176 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25176.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haoming Wen, Yushi Bai, Juanzi Li, Jie Tang</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> We introduce SIRI, Scaling Iterative Reinforcement Learning with Interleaved Compression, a simple yet effective RL approach for Large Reasoning Models (LRMs) that enables more efficient and accurate reasoning. Existing studies have observed repetitive thinking patterns in LRMs, and attempts to reduce them often come at the cost of performance. In this paper, we show that this trade-off can be overcome through a training regime that iteratively alternates between compressing and expanding the reasoning budget, by dynamically adjusting the maximum rollout length during training. The compression phase cuts the rollout length, forcing the model to make precise and valuable decisions within a limited context, which effectively reduces redundant tokens and increases reasoning density. The expansion phase then relaxes the length limit, providing space for the model to explore and plan in long-horizon settings. Remarkably, we find that after each compression-expansion cycle, the model's performance improves even as its output length decreases, steadily pushing it closer to the Pareto frontier in the performance-efficiency trade-off. Training on DeepSeek-R1-Distill-Qwen-1.5B, SIRI-low improves performance on AIME24 by 43.2% while reducing token usage by 46.9% after three iterations, and SIRI-high achieves the highest accuracy compared to all other methods (Figure 1). Our findings shed light on the potential of periodically oscillating the LRM's output truncation length during training to dynamically balance exploration and efficiency in reasoning, converging towards an optimal "sweet spot" between the two. Our models are publicly available.</p> - https://arxiv.org/abs/2509.25176 - Mon, 29 Sep 2025 17:59:08 +0000 - - - PixelCraft: A Multi-Agent System for High-Fidelity Visual Reasoning on Structured Images - https://arxiv.org/abs/2509.25185 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25185.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shuoshuo Zhang, Zijian Li, Yizhen Zhang, Jingjing Fu, Lei Song, Jiang Bian, Jun Zhang, Yujiu Yang, Rui Wang</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Structured images (e.g., charts and geometric diagrams) remain challenging for multimodal large language models (MLLMs), as perceptual slips can cascade into erroneous conclusions. Intermediate visual cues can steer reasoning; however, existing cue-based methods are constrained with low-fidelity image processing and linear, rigid reasoning patterns, limiting their effectiveness on complex structured-image tasks. In this paper, we propose PixelCraft, a novel multi-agent system for high-fidelity image processing and flexible visual reasoning on structured images. The system comprises a dispatcher, a planner, a reasoner, critics, and a set of visual tool agents. To achieve high-fidelity processing, we construct a high-quality corpus and fine-tune an MLLM into a grounding model, whose pixel-level localizations are integrated with traditional computer vision (CV) algorithms in tool agents. Building on this foundation, PixelCraft facilitates flexible visual reasoning through a dynamic three-stage workflow of tool selection, agent discussion, and self-criticism. Moreover, unlike prior linear reasoning patterns that simply append historical images, PixelCraft maintains an image memory to allow the planner to adaptively revisit earlier visual steps, explore alternative reasoning branches, and dynamically adjust the reasoning trajectory during discussion. Extensive experiments on challenging chart and geometry benchmarks demonstrate that PixelCraft significantly improves visual reasoning performance for advanced MLLMs, setting a new standard for structured image reasoning. Our code will be available at https://github.com/microsoft/PixelCraft.</p> - https://arxiv.org/abs/2509.25185 - Mon, 29 Sep 2025 17:59:49 +0000 - - - Visual Jigsaw Post-Training Improves MLLMs - https://arxiv.org/abs/2509.25190 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25190.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Penghao Wu, Yushan Zhang, Haiwen Diao, Bo Li, Lewei Lu, Ziwei Liu</p><p><b>Upvotes:</b> 33</p><p><b>Summary:</b> Reinforcement learning based post-training has recently emerged as a powerful paradigm for enhancing the alignment and reasoning capabilities of multimodal large language models (MLLMs). While vision-centric post-training is crucial for enhancing MLLMs' intrinsic understanding of visual signals, current post-training paradigms are predominantly text-centric, where dense visual inputs are only leveraged to extract sparse cues for text-based reasoning. There exist a few approaches in this direction, however, they often still rely on text as an intermediate mediator or introduce additional visual generative designs. In this work, we introduce Visual Jigsaw, a generic self-supervised post-training framework designed to strengthen visual understanding in MLLMs. Visual Jigsaw is formulated as a general ordering task: visual inputs are partitioned, shuffled, and the model must reconstruct the visual information by producing the correct permutation in natural language. This naturally aligns with reinforcement learning from verifiable rewards (RLVR), requires no additional visual generative components, and derives its supervisory signal automatically without any annotations. We instantiate Visual Jigsaw across three visual modalities, including images, videos, and 3D data. Extensive experiments demonstrate substantial improvements in fine-grained perception, temporal reasoning, and 3D spatial understanding. Our findings highlight the potential of self-supervised vision-centric tasks in post-training MLLMs and aim to inspire further research on vision-centric pretext designs. Project Page: https://penghao-wu.github.io/visual_jigsaw/</p> - https://arxiv.org/abs/2509.25190 + InfoAgent: Advancing Autonomous Information-Seeking Agents + https://arxiv.org/abs/2509.25189 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25189.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Gongrui Zhang, Jialiang Zhu, Ruiqi Yang, Kai Qiu, Miaosen Zhang, Zhirong Wu, Qi Dai, Bei Liu, Chong Luo, Zhengyuan Yang, Linjie Li, Lijuan Wang, Weizhu Chen, Yuan Zhang, Xin Li, Zhaoyi Liu, Xin Geng, Baining Guo</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> Building Large Language Model agents that expand their capabilities by interacting with external tools represents a new frontier in AI research and applications. In this paper, we introduce InfoAgent, a deep research agent powered by an innovative data synthesis pipeline and orchestrated web search tools. To construct challenging, hard-to-find queries,we build entity trees and apply sub-tree sampling with entity fuzzification to systematically increase question difficulty. Unlike prior work that relies heavily on commercial search tools, we develop a dedicated self-hosted search infrastructure, enhancing transparency of agent environments and facilitating further advancement of agent capacity. We evaluate the effectiveness of our data pipeline by measuring the average number of tool calls required to correctly answer a question, and also show that our agent yields better performance when equipped with our tools. Our InfoAgent is post-trained from Qwen3-14B using a two-stage recipe: cold-start supervised finetuning to instill long-horizon search behaviors, followed by reinforcement learning which significantly improves reasoning-driven tool use. With our methods, InfoAgent achieves 15.3\% accuracy on BrowseComp, 29.2\% on BrowseComp-ZH, and 40.4\% on Xbench-DS, outperforming prior open-source deep research agents such as WebSailor-72B and DeepDive-32B.</p> + https://arxiv.org/abs/2509.25189 Mon, 29 Sep 2025 17:59:57 +0000 - VGGT-X: When VGGT Meets Dense Novel View Synthesis - https://arxiv.org/abs/2509.25191 - <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25191.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yang Liu, Chuanchen Luo, Zimo Tang, Junran Peng, Zhaoxiang Zhang</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> We study the problem of applying 3D Foundation Models (3DFMs) to dense Novel View Synthesis (NVS). Despite significant progress in Novel View Synthesis powered by NeRF and 3DGS, current approaches remain reliant on accurate 3D attributes (e.g., camera poses and point clouds) acquired from Structure-from-Motion (SfM), which is often slow and fragile in low-texture or low-overlap captures. Recent 3DFMs showcase orders of magnitude speedup over the traditional pipeline and great potential for online NVS. But most of the validation and conclusions are confined to sparse-view settings. Our study reveals that naively scaling 3DFMs to dense views encounters two fundamental barriers: dramatically increasing VRAM burden and imperfect outputs that degrade initialization-sensitive 3D training. To address these barriers, we introduce VGGT-X, incorporating a memory-efficient VGGT implementation that scales to 1,000+ images, an adaptive global alignment for VGGT output enhancement, and robust 3DGS training practices. Extensive experiments show that these measures substantially close the fidelity gap with COLMAP-initialized pipelines, achieving state-of-the-art results in dense COLMAP-free NVS and pose estimation. Additionally, we analyze the causes of remaining gaps with COLMAP-initialized rendering, providing insights for the future development of 3D foundation models and dense NVS. Our project page is available at https://dekuliutesla.github.io/vggt-x.github.io/</p> - https://arxiv.org/abs/2509.25191 - Mon, 29 Sep 2025 17:59:59 +0000 + VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes + https://arxiv.org/abs/2509.25339 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25339.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Paul Gavrikov, Wei Lin, M. Jehanzeb Mirza, Soumya Jahagirdar, Muhammad Huzaifa, Sivan Doveh, Serena Yeung-Levy, James Glass, Hilde Kuehne</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Is basic visual understanding really solved in state-of-the-art VLMs? We present VisualOverload, a slightly different visual question answering (VQA) benchmark comprising 2,720 question-answer pairs, with privately held ground-truth responses. Unlike prior VQA datasets that typically focus on near global image understanding, VisualOverload challenges models to perform simple, knowledge-free vision tasks in densely populated (or, overloaded) scenes. Our dataset consists of high-resolution scans of public-domain paintings that are populated with multiple figures, actions, and unfolding subplots set against elaborately detailed backdrops. We manually annotated these images with questions across six task categories to probe for a thorough understanding of the scene. We hypothesize that current benchmarks overestimate the performance of VLMs, and encoding and reasoning over details is still a challenging task for them, especially if they are confronted with densely populated scenes. Indeed, we observe that even the best model (o3) out of 37 tested models only achieves 19.6% accuracy on our hardest test split and overall 69.5% accuracy on all questions. Beyond a thorough evaluation, we complement our benchmark with an error analysis that reveals multiple failure modes, including a lack of counting skills, failure in OCR, and striking logical inconsistencies under complex tasks. Altogether, VisualOverload exposes a critical gap in current vision models and offers a crucial resource for the community to develop better models. Benchmark: http://paulgavrikov.github.io/visualoverload</p> + https://arxiv.org/abs/2509.25339 + Mon, 29 Sep 2025 18:00:25 +0000 + + + A Cartography of Open Collaboration in Open Source AI: Mapping Practices, Motivations, and Governance in 14 Open Large Language Model Projects + https://arxiv.org/abs/2509.25397 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25397.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Johan Linåker, Cailean Osborne, Jennifer Ding, Ben Burtenshaw</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> The proliferation of open large language models (LLMs) is fostering a vibrant ecosystem of research and innovation in artificial intelligence (AI). However, the methods of collaboration used to develop open LLMs both before and after their public release have not yet been comprehensively studied, limiting our understanding of how open LLM projects are initiated, organized, and governed as well as what opportunities there are to foster this ecosystem even further. We address this gap through an exploratory analysis of open collaboration throughout the development and reuse lifecycle of open LLMs, drawing on semi-structured interviews with the developers of 14 open LLMs from grassroots projects, research institutes, startups, and Big Tech companies in North America, Europe, Africa, and Asia. We make three key contributions to research and practice. First, collaboration in open LLM projects extends far beyond the LLMs themselves, encompassing datasets, benchmarks, open source frameworks, leaderboards, knowledge sharing and discussion forums, and compute partnerships, among others. Second, open LLM developers have a variety of social, economic, and technological motivations, from democratizing AI access and promoting open science to building regional ecosystems and expanding language representation. Third, the sampled open LLM projects exhibit five distinct organizational models, ranging from single company projects to non-profit-sponsored grassroots projects, which vary in their centralization of control and community engagement strategies used throughout the open LLM lifecycle. We conclude with practical recommendations for stakeholders seeking to support the global community building a more open future for AI.</p> + https://arxiv.org/abs/2509.25397 + Mon, 29 Sep 2025 18:55:18 +0000 + + + Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play + https://arxiv.org/abs/2509.25541 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25541.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Qinsi Wang, Bo Liu, Tianyi Zhou, Jing Shi, Yueqian Lin, Yiran Chen, Hai Helen Li, Kun Wan, Wentian Zhao</p><p><b>Upvotes:</b> 100</p><p><b>Summary:</b> Although reinforcement learning (RL) can effectively enhance the reasoning capabilities of vision-language models (VLMs), current methods remain heavily dependent on labor-intensive datasets that require extensive manual construction and verification, leading to extremely high training costs and consequently constraining the practical deployment of VLMs. To address this challenge, we propose Vision-Zero, a domain-agnostic framework enabling VLM self-improvement through competitive visual games generated from arbitrary image pairs. Specifically, Vision-Zero encompasses three main attributes: (1) Strategic Self-Play Framework: Vision-Zero trains VLMs in "Who Is the Spy"-style games, where the models engage in strategic reasoning and actions across multiple roles. Through interactive gameplay, models autonomously generate their training data without human annotation. (2) Gameplay from Arbitrary Images: Unlike existing gamified frameworks, Vision-Zero can generate games from arbitrary images, thereby enhancing the model's reasoning ability across diverse domains and showing strong generalization to different tasks. We demonstrate this versatility using three distinct types of image datasets: CLEVR-based synthetic scenes, charts, and real-world images. (3) Sustainable Performance Gain: We introduce Iterative Self-Play Policy Optimization (Iterative-SPO), a novel training algorithm that alternates between Self-Play and reinforcement learning with verifiable rewards (RLVR), mitigating the performance plateau often seen in self-play-only training and achieving sustained long-term improvements. Despite using label-free data, Vision-Zero achieves state-of-the-art performance on reasoning, chart question answering, and vision-centric understanding tasks, surpassing other annotation-based methods. Models and code has been released at https://github.com/wangqinsi1/Vision-Zero.</p> + https://arxiv.org/abs/2509.25541 + Mon, 29 Sep 2025 21:55:55 +0000 + + + Swift: An Autoregressive Consistency Model for Efficient Weather Forecasting + https://arxiv.org/abs/2509.25631 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25631.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jason Stock, Troy Arcomano, Rao Kotamarthi</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Diffusion models offer a physically grounded framework for probabilistic weather forecasting, but their typical reliance on slow, iterative solvers during inference makes them impractical for subseasonal-to-seasonal (S2S) applications where long lead-times and domain-driven calibration are essential. To address this, we introduce Swift, a single-step consistency model that, for the first time, enables autoregressive finetuning of a probability flow model with a continuous ranked probability score (CRPS) objective. This eliminates the need for multi-model ensembling or parameter perturbations. Results show that Swift produces skillful 6-hourly forecasts that remain stable for up to 75 days, running 39times faster than state-of-the-art diffusion baselines while achieving forecast skill competitive with the numerical-based, operational IFS ENS. This marks a step toward efficient and reliable ensemble forecasting from medium-range to seasonal-scales.</p> + https://arxiv.org/abs/2509.25631 + Tue, 30 Sep 2025 00:54:24 +0000 + + + Nudging the Boundaries of LLM Reasoning + https://arxiv.org/abs/2509.25666 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25666.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Justin Chih-Yao Chen, Becky Xiangyu Peng, Prafulla Kumar Choubey, Kung-Hsiang Huang, Jiaxin Zhang, Mohit Bansal, Chien-Sheng Wu</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Current online reinforcement learning (RL) algorithms like GRPO share a key limitation in LLM reasoning: they cannot learn from problems that are "unsolvable" to the model. In other words, they can only improve performance on problems where the model is capable of exploring the correct answer. Consequently, the model's "upper limit" remains unchanged after RL training, even though the likelihood of solving easier, solvable problems may increase. These hard samples cannot contribute to training, as no rollouts yield rewards and thus no gradients are produced. To unlock learning from these hard samples, we propose NuRL, a "nudging" method that aims to push the upper bound of LLM reasoning using self-generated hints, i.e., abstract cues that help reduce the problem difficulty for the model. Given a question and its gold answer, the model generates a CoT and then produces a hint containing the core knowledge needed to solve the problem. During training, we generate G rollouts from the base policy and use the pass rate to decide whether the hint should be injected. For hard samples with a 0% pass rate, we inject the hint and regenerate a new batch of trajectories. This yields two benefits: (1) the hint boosts pass rates (from 0% to non-zero), thereby introducing training signals for previously unsolvable samples, and (2) the hints are self-generated, avoiding distributional shift and do not rely on external models. NuRL achieves consistent improvements across 6 benchmarks and 3 models, while remaining complementary to test-time scaling. Notably, NuRL can raise the model's upper limit, whereas GRPO leaves pass@1024 unchanged from the base model. Furthermore, we present a systematic study of what makes an effective hint and when hints are most useful. Interestingly, the best hints are abstract and high-level, and are most beneficial when applied necessarily and after GRPO has converged.</p> + https://arxiv.org/abs/2509.25666 + Tue, 30 Sep 2025 02:01:40 +0000 + + + DeepCodeSeek: Real-Time API Retrieval for Context-Aware Code Generation + https://arxiv.org/abs/2509.25716 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25716.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Esakkivel Esakkiraja, Denis Akhiyarov, Aditya Shanmugham, Chitra Ganapathy</p><p><b>Upvotes:</b> 3</p><p><b>Summary:</b> Current search techniques are limited to standard RAG query-document applications. In this paper, we propose a novel technique to expand the code and index for predicting the required APIs, directly enabling high-quality, end-to-end code generation for auto-completion and agentic AI applications. We address the problem of API leaks in current code-to-code benchmark datasets by introducing a new dataset built from real-world ServiceNow Script Includes that capture the challenge of unclear API usage intent in the code. Our evaluation metrics show that this method achieves 87.86% top-40 retrieval accuracy, allowing the critical context with APIs needed for successful downstream code generation. To enable real-time predictions, we develop a comprehensive post-training pipeline that optimizes a compact 0.6B reranker through synthetic dataset generation, supervised fine-tuning, and reinforcement learning. This approach enables our compact reranker to outperform a much larger 8B model while maintaining 2.5x reduced latency, effectively addressing the nuances of enterprise-specific code without the computational overhead of larger models.</p> + https://arxiv.org/abs/2509.25716 + Tue, 30 Sep 2025 03:23:27 +0000 + + + Thinking Sparks!: Emergent Attention Heads in Reasoning Models During Post Training + https://arxiv.org/abs/2509.25758 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25758.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yein Park, Minbyul Jeong, Jaewoo Kang</p><p><b>Upvotes:</b> 18</p><p><b>Summary:</b> The remarkable capabilities of modern large reasoning models are largely unlocked through post-training techniques such as supervised fine-tuning and reinforcement learning. However, the architectural mechanisms behind such improvements remain largely opaque. In this work, we use circuit analysis to demonstrate that post-training for complex reasoning sparks the emergence of novel, functionally specialized attention heads. These heads collectively support structured reasoning and computation. Our comparative analysis across Qwen families and DeepSeek-distilled model reveals that these emergent heads evolve differently under different training regimes. Distillation and SFT foster a cumulative addition of stable reasoning heads. In contrast, group relative policy optimization operates in a dynamic search mode: relatively few attention heads are iteratively activated, evaluated, and pruned, with their survival closely tracking fluctuations in the task reward signal. Furthermore, we find that controllable think on/off models do not possess dedicated thinking heads. Instead, turning off explicit reasoning triggers a broader-but less efficient-set of compensatory heads. Through ablation and qualitative analyses, we connect these circuit-level dynamics to a crucial performance trade-off: strengthened heads enable sophisticated problem-solving strategies for difficult problems but can also introduce over-thinking failure modes, such as calculation errors or logical loops on simpler tasks. These findings connect circuit-level dynamics to macro-level performance, identifying an inherent tension where complex reasoning comes at the cost of elementary computations. More broadly, our work points to future directions for training policy design, emphasizing the need to balance the development of effective reasoning strategies with the assurance of reliable, flawless execution.</p> + https://arxiv.org/abs/2509.25758 + Tue, 30 Sep 2025 04:23:43 +0000 + + + TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning + https://arxiv.org/abs/2509.25760 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25760.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhepei Wei, Xiao Yang, Kai Sun, Jiaqi Wang, Rulin Shao, Sean Chen, Mohammad Kachuee, Teja Gollapudi, Tony Liao, Nicolas Scheffer, Rakesh Wanga, Anuj Kumar, Yu Meng, Wen-tau Yih, Xin Luna Dong</p><p><b>Upvotes:</b> 45</p><p><b>Summary:</b> While large language models (LLMs) have demonstrated strong performance on factoid question answering, they are still prone to hallucination and untruthful responses, particularly when tasks demand information outside their parametric knowledge. Indeed, truthfulness requires more than accuracy -- models must also recognize uncertainty and abstain when unsure to avoid hallucinations. This presents a fundamental challenge for existing methods: approaches that optimize for accuracy often amplify hallucinations, while those that encourage abstention can become overly conservative, sacrificing correct answers. Both extremes ultimately compromise truthfulness. In this work, we present TruthRL, a general reinforcement learning (RL) framework that directly optimizes the truthfulness of LLMs. Specifically, we implement TruthRL using GRPO with a simple yet effective ternary reward that distinguishes correct answers, hallucinations, and abstentions. It incentivizes models to reduce hallucinations not only by providing correct responses, but also by enabling abstention when uncertain, thereby improving truthfulness. Extensive experiments across four knowledge-intensive benchmarks show that, compared to vanilla RL, TruthRL significantly reduces hallucinations by 28.9% and improves truthfulness by 21.1%, with consistent gains across various backbone models (e.g., Qwen, Llama) under both retrieval and non-retrieval setups. In-depth ablation study demonstrates that vanilla accuracy-driven methods, such as supervised fine-tuning or RL with a binary reward, struggle to balance factual correctness and uncertainty. In contrast, our proposed truthfulness-driven TruthRL achieves strong performance in both accuracy and truthfulness, underscoring the importance of learning objective design for developing truthful LLMs.</p> + https://arxiv.org/abs/2509.25760 + Tue, 30 Sep 2025 04:25:17 +0000 + + + Learning to Reason as Action Abstractions with Scalable Mid-Training RL + https://arxiv.org/abs/2509.25810 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25810.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shenao Zhang, Donghan Yu, Yihao Feng, Bowen Jin, Zhaoran Wang, John Peebles, Zirui Wang</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Large language models excel with reinforcement learning (RL), but fully unlocking this potential requires a mid-training stage. An effective mid-training phase should identify a compact set of useful actions and enable fast selection among them through online RL. We formalize this intuition by presenting the first theoretical result on how mid-training shapes post-training: it characterizes an action subspace that minimizes both the value approximation error from pruning and the RL error during subsequent planning. Our analysis reveals two key determinants of mid-training effectiveness: pruning efficiency, which shapes the prior of the initial RL policy, and its impact on RL convergence, which governs the extent to which that policy can be improved via online interactions. These results suggest that mid-training is most effective when the decision space is compact and the effective horizon is short, highlighting the importance of operating in the space of action abstractions rather than primitive actions. Building on these insights, we propose Reasoning as Action Abstractions (RA3), a scalable mid-training algorithm. Specifically, we derive a sequential variational lower bound and optimize it by iteratively discovering temporally-consistent latent structures via RL, followed by fine-tuning on the bootstrapped data. Experiments on code generation tasks demonstrate the effectiveness of our approach. Across multiple base models, RA3 improves the average performance on HumanEval and MBPP by 8 and 4 points over the base model and the next-token prediction baseline. Furthermore, RA3 achieves faster convergence and higher asymptotic performance in RLVR on HumanEval+, MBPP+, LiveCodeBench, and Codeforces.</p> + https://arxiv.org/abs/2509.25810 + Tue, 30 Sep 2025 05:34:20 +0000 + + + More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models + https://arxiv.org/abs/2509.25848 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25848.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xinyu Tian, Shu Zou, Zhaoyuan Yang, Mengqi He, Fabian Waschkowski, Lukas Wesemann, Peter Tu, Jing Zhang</p><p><b>Upvotes:</b> 28</p><p><b>Summary:</b> Reasoning has emerged as a pivotal capability in Large Language Models (LLMs). Through Reinforcement Learning (RL), typically Group Relative Policy Optimization (GRPO), these models are able to solve complex tasks such as mathematics and code generation. Building on these advances, recent research has sought to extend reasoning to Vision-Language Models (VLMs), yielding promising results across diverse visual tasks. Despite this progress, our study uncovers the dual nature of multimodal reasoning: while it substantially enhances logical inference and facilitates performance on challenging problems, it may gradually impair perceptual grounding, leading to recognition failures on otherwise basic visual questions. Through further analysis, we attribute this phenomenon to visual forgetting, wherein prolonged reasoning causes the model to increasingly disregard visual input. To address this, we propose Vision-Anchored Policy Optimization (VAPO), a simple yet effective method that explicitly steers the reasoning process toward visually grounded trajectories. Our result model, VAPO-Thinker-7B, significantly strengthens the model's reliance on visual information and achieves new state-of-the-art results on a wide range of established benchmarks. Project page: https://xytian1008.github.io/VAPO/</p> + https://arxiv.org/abs/2509.25848 + Tue, 30 Sep 2025 06:37:47 +0000 + + + Mem-α: Learning Memory Construction via Reinforcement Learning + https://arxiv.org/abs/2509.25911 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.25911.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yu Wang, Ryuichi Takanobu, Zhiqi Liang, Yuzhen Mao, Yuanzhe Hu, Julian McAuley, Xiaojian Wu</p><p><b>Upvotes:</b> 10</p><p><b>Summary:</b> Large language model (LLM) agents are constrained by limited context windows, necessitating external memory systems for long-term information understanding. Current memory-augmented agents typically depend on pre-defined instructions and tools for memory updates. However, language models may lack the ability to determine which information to store, how to structure it, and when to update it, especially as memory systems become more complex. This results in suboptimal memory construction and information loss. To this end, we propose Mem-alpha, a reinforcement learning framework that trains agents to effectively manage complex memory systems through interaction and feedback. We also construct a specialized training dataset spanning diverse multi-turn interaction patterns paired with comprehensive evaluation questions designed to teach effective memory management. During training, agents process sequential information chunks, learn to extract and store relevant content, then update the memory system. The reward signal derives from downstream question-answering accuracy over the full interaction history, directly optimizing for memory construction. To illustrate the effectiveness of our training framework, we design a memory architecture comprising core, episodic, and semantic components, equipped with multiple tools for memory operations. Empirical evaluation demonstrates that Mem-alpha achieves significant improvements over existing memory-augmented agent baselines. Despite being trained exclusively on instances with a maximum length of 30k tokens, our agents exhibit remarkable generalization to sequences exceeding 400k tokens, over 13x the training length, highlighting the robustness of Mem-alpha.</p> + https://arxiv.org/abs/2509.25911 + Tue, 30 Sep 2025 08:02:34 +0000 + + + Muon Outperforms Adam in Tail-End Associative Memory Learning + https://arxiv.org/abs/2509.26030 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26030.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Shuche Wang, Fengzhuo Zhang, Jiaxiang Li, Cunxiao Du, Chao Du, Tianyu Pang, Zhuoran Yang, Mingyi Hong, Vincent Y. F. Tan</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> The Muon optimizer is consistently faster than Adam in training Large Language Models (LLMs), yet the mechanism underlying its success remains unclear. This paper demystifies this mechanism through the lens of associative memory. By ablating the transformer components optimized by Muon, we reveal that the associative memory parameters of LLMs, namely the Value and Output (VO) attention weights and Feed-Forward Networks (FFNs), are the primary contributors to Muon's superiority. Motivated by this associative memory view, we then explain Muon's superiority on real-world corpora, which are intrinsically heavy-tailed: a few classes (tail classes) appear far less frequently than others. The superiority is explained through two key properties: (i) its update rule consistently yields a more isotropic singular spectrum than Adam; and as a result, (ii) on heavy-tailed data, it optimizes tail classes more effectively than Adam. Beyond empirical evidence, we theoretically confirm these findings by analyzing a one-layer associative memory model under class-imbalanced data. We prove that Muon consistently achieves balanced learning across classes regardless of feature embeddings, whereas Adam can induce large disparities in learning errors depending on embedding properties. In summary, our empirical observations and theoretical analyses reveal Muon's core advantage: its update rule aligns with the outer-product structure of linear associative memories, enabling more balanced and effective learning of tail classes in heavy-tailed distributions than Adam.</p> + https://arxiv.org/abs/2509.26030 + Tue, 30 Sep 2025 10:04:08 +0000 + + + EntroPE: Entropy-Guided Dynamic Patch Encoder for Time Series Forecasting + https://arxiv.org/abs/2509.26157 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26157.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Transformer-based models have significantly advanced time series forecasting, with patch-based input strategies offering efficiency and improved long-horizon modeling. Yet, existing approaches rely on temporally-agnostic patch construction, where arbitrary starting positions and fixed lengths fracture temporal coherence by splitting natural transitions across boundaries. This naive segmentation often disrupts short-term dependencies and weakens representation learning. In response, we propose EntroPE (Entropy-Guided Dynamic Patch Encoder), a novel, temporally informed framework that dynamically detects transition points via conditional entropy and dynamically places patch boundaries. This preserves temporal structure while retaining the computational benefits of patching. EntroPE consists of two key modules, namely an Entropy-based Dynamic Patcher (EDP) that applies information-theoretic criteria to locate natural temporal shifts and determine patch boundaries, and an Adaptive Patch Encoder (APE) that employs pooling and cross-attention to capture intra-patch dependencies and produce fixed-size latent representations. These embeddings are then processed by a global transformer to model inter-patch dynamics. Experiments across long-term forecasting benchmarks demonstrate that EntroPE improves both accuracy and efficiency, establishing entropy-guided dynamic patching as a promising new paradigm for time series modeling. Code is available at: https://github.com/Sachithx/EntroPE.</p> + https://arxiv.org/abs/2509.26157 + Tue, 30 Sep 2025 12:09:56 +0000 + + + Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners + https://arxiv.org/abs/2509.26226 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26226.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xin Xu, Cliveb AI, Kai Yang, Tianhao Chen, Yang Wang, Saiyong Yang, Can Yang</p><p><b>Upvotes:</b> 26</p><p><b>Summary:</b> Reinforcement Learning with Verifiable Reward (RLVR) effectively solves complex tasks but demands extremely long context lengths during training, leading to substantial computational costs. While multi-stage training can partially mitigate this, starting with overly short contexts often causes irreversible performance degradation, ultimately failing to reduce overall training compute significantly. In this paper, we introduce **T**hinking-**F**ree **P**olicy **I**nitialization (**TFPI**), a simple yet effective adaptation to RLVR that bridges long Chain-of-Thought (CoT) distillation and standard RLVR. TFPI employs a simple *ThinkFree* operation, explicitly discarding the thinking content via a direct *</think>* append, to reduce token usage during inference. Training with *ThinkFree*-adapted inputs improves performance and lowers token consumption, even in the original slow-thinking mode. Extensive experiments across various benchmarks have shown that TFPI accelerates RL convergence, achieves a higher performance ceiling, and yields more token-efficient reasoning models without specialized rewards or complex training designs. With TFPI only, we train a 4B model to reach 89.0% accuracy on AIME24 and 65.5% on LiveCodeBench using less than 4K H20 hours.</p> + https://arxiv.org/abs/2509.26226 + Tue, 30 Sep 2025 13:25:00 +0000 + + + IMG: Calibrating Diffusion Models via Implicit Multimodal Guidance + https://arxiv.org/abs/2509.26231 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26231.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jiayi Guo, Chuanhao Yan, Xingqian Xu, Yulin Wang, Kai Wang, Gao Huang, Humphrey Shi</p><p><b>Upvotes:</b> 13</p><p><b>Summary:</b> Ensuring precise multimodal alignment between diffusion-generated images and input prompts has been a long-standing challenge. Earlier works finetune diffusion weight using high-quality preference data, which tends to be limited and difficult to scale up. Recent editing-based methods further refine local regions of generated images but may compromise overall image quality. In this work, we propose Implicit Multimodal Guidance (IMG), a novel re-generation-based multimodal alignment framework that requires no extra data or editing operations. Specifically, given a generated image and its prompt, IMG a) utilizes a multimodal large language model (MLLM) to identify misalignments; b) introduces an Implicit Aligner that manipulates diffusion conditioning features to reduce misalignments and enable re-generation; and c) formulates the re-alignment goal into a trainable objective, namely Iteratively Updated Preference Objective. Extensive qualitative and quantitative evaluations on SDXL, SDXL-DPO, and FLUX show that IMG outperforms existing alignment methods. Furthermore, IMG acts as a flexible plug-and-play adapter, seamlessly enhancing prior finetuning-based alignment methods. Our code will be available at https://github.com/SHI-Labs/IMG-Multimodal-Diffusion-Alignment.</p> + https://arxiv.org/abs/2509.26231 + Tue, 30 Sep 2025 13:27:03 +0000 + + + ProfVLM: A Lightweight Video-Language Model for Multi-View Proficiency Estimation + https://arxiv.org/abs/2509.26278 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26278.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Edoardo Bianchi, Jacopo Staiano, Antonio Liotta</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Existing approaches to skill proficiency estimation often rely on black-box video classifiers, ignoring multi-view context and lacking explainability. We present ProfVLM, a compact vision-language model that reformulates this task as generative reasoning: it jointly predicts skill level and generates expert-like feedback from egocentric and exocentric videos. Central to our method is an AttentiveGatedProjector that dynamically fuses multi-view features, projected from a frozen TimeSformer backbone into a language model tuned for feedback generation. Trained on EgoExo4D with expert commentaries, ProfVLM surpasses state-of-the-art methods while using up to 20x fewer parameters and reducing training time by up to 60%. Our approach not only achieves superior accuracy across diverse activities, but also outputs natural language critiques aligned with performance, offering transparent reasoning. These results highlight generative vision-language modeling as a powerful new direction for skill assessment.</p> + https://arxiv.org/abs/2509.26278 + Tue, 30 Sep 2025 14:00:41 +0000 + + + TAU: A Benchmark for Cultural Sound Understanding Beyond Semantics + https://arxiv.org/abs/2509.26329 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26329.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yi-Cheng Lin, Yu-Hua Chen, Jia-Kai Dong, Yueh-Hsuan Huang, Szu-Chi Chen, Yu-Chen Chen, Chih-Yao Chen, Yu-Jung Lin, Yu-Ling Chen, Zih-Yu Chen, I-Ning Tsai, Hsiu-Hsuan Wang, Ho-Lam Chung, Ke-Han Lu, Hung-yi Lee</p><p><b>Upvotes:</b> 2</p><p><b>Summary:</b> Large audio-language models are advancing rapidly, yet most evaluations emphasize speech or globally sourced sounds, overlooking culturally distinctive cues. This gap raises a critical question: can current models generalize to localized, non-semantic audio that communities instantly recognize but outsiders do not? To address this, we present TAU (Taiwan Audio Understanding), a benchmark of everyday Taiwanese "soundmarks." TAU is built through a pipeline combining curated sources, human editing, and LLM-assisted question generation, producing 702 clips and 1,794 multiple-choice items that cannot be solved by transcripts alone. Experiments show that state-of-the-art LALMs, including Gemini 2.5 and Qwen2-Audio, perform far below local humans. TAU demonstrates the need for localized benchmarks to reveal cultural blind spots, guide more equitable multimodal evaluation, and ensure models serve communities beyond the global mainstream.</p> + https://arxiv.org/abs/2509.26329 + Tue, 30 Sep 2025 14:40:45 +0000 + + + MotionRAG: Motion Retrieval-Augmented Image-to-Video Generation + https://arxiv.org/abs/2509.26391 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26391.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Chenhui Zhu, Yilu Wu, Shuai Wang, Gangshan Wu, Limin Wang</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> Image-to-video generation has made remarkable progress with the advancements in diffusion models, yet generating videos with realistic motion remains highly challenging. This difficulty arises from the complexity of accurately modeling motion, which involves capturing physical constraints, object interactions, and domain-specific dynamics that are not easily generalized across diverse scenarios. To address this, we propose MotionRAG, a retrieval-augmented framework that enhances motion realism by adapting motion priors from relevant reference videos through Context-Aware Motion Adaptation (CAMA). The key technical innovations include: (i) a retrieval-based pipeline extracting high-level motion features using video encoder and specialized resamplers to distill semantic motion representations; (ii) an in-context learning approach for motion adaptation implemented through a causal transformer architecture; (iii) an attention-based motion injection adapter that seamlessly integrates transferred motion features into pretrained video diffusion models. Extensive experiments demonstrate that our method achieves significant improvements across multiple domains and various base models, all with negligible computational overhead during inference. Furthermore, our modular design enables zero-shot generalization to new domains by simply updating the retrieval database without retraining any components. This research enhances the core capability of video generation systems by enabling the effective retrieval and transfer of motion priors, facilitating the synthesis of realistic motion dynamics.</p> + https://arxiv.org/abs/2509.26391 + Tue, 30 Sep 2025 15:26:04 +0000 + + + Regression Language Models for Code + https://arxiv.org/abs/2509.26476 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26476.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yash Akhauri, Xingyou Song, Arissa Wongpanich, Bryan Lewandowski, Mohamed S. Abdelfattah</p><p><b>Upvotes:</b> 6</p><p><b>Summary:</b> We study code-to-metric regression: predicting numeric outcomes of code executions, a challenging task due to the open-ended nature of programming languages. While prior methods have resorted to heavy and domain-specific feature engineering, we show that a single unified Regression Language Model (RLM) can simultaneously predict directly from text, (i) the memory footprint of code across multiple high-level languages such as Python and C++, (ii) the latency of Triton GPU kernels, and (iii) the accuracy and speed of trained neural networks represented in ONNX. In particular, a relatively small 300M parameter RLM initialized from T5Gemma, obtains > 0.9 Spearman-rank on competitive programming submissions from APPS, and a single unified model achieves > 0.5 average Spearman-rank across 17 separate languages from CodeNet. Furthermore, the RLM can obtain the highest average Kendall-Tau of 0.46 on five classic NAS design spaces previously dominated by graph neural networks, and simultaneously predict architecture latencies on numerous hardware platforms.</p> + https://arxiv.org/abs/2509.26476 + Tue, 30 Sep 2025 16:25:23 +0000 + + + dParallel: Learnable Parallel Decoding for dLLMs + https://arxiv.org/abs/2509.26488 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26488.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu, Xinchao Wang</p><p><b>Upvotes:</b> 16</p><p><b>Summary:</b> Diffusion large language models (dLLMs) have recently drawn considerable attention within the research community as a promising alternative to autoregressive generation, offering parallel token prediction and lower inference latency. Yet, their parallel decoding potential remains largely underexplored, as existing open-source models still require nearly token-length decoding steps to ensure performance. To address this, we introduce dParallel, a simple and effective method that unlocks the inherent parallelism of dLLMs for fast sampling. We identify that the key bottleneck to parallel decoding arises from the sequential certainty convergence for masked tokens. Building on this insight, we introduce the core of our approach: certainty-forcing distillation, a novel training strategy that distills the model to follow its original sampling trajectories while enforcing it to achieve high certainty on masked tokens more rapidly and in parallel. Extensive experiments across various benchmarks demonstrate that our method can dramatically reduce the number of decoding steps while maintaining performance. When applied to the LLaDA-8B-Instruct model, dParallel reduces decoding steps from 256 to 30 on GSM8K, achieving an 8.5x speedup without performance degradation. On the MBPP benchmark, it cuts decoding steps from 256 to 24, resulting in a 10.5x speedup while maintaining accuracy. Our code is available at https://github.com/czg1225/dParallel</p> + https://arxiv.org/abs/2509.26488 + Tue, 30 Sep 2025 16:32:52 +0000 + + + VitaBench: Benchmarking LLM Agents with Versatile Interactive Tasks in Real-world Applications + https://arxiv.org/abs/2509.26490 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26490.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Wei He, Yueqing Sun, Hongyan Hao, Xueyuan Hao, Zhikang Xia, Qi Gu, Chengcheng Han, Dengchang Zhao, Hui Su, Kefeng Zhang, Man Gao, Xi Su, Xiaodong Cai, Xunliang Cai, Yu Yang, Yunke Zhao</p><p><b>Upvotes:</b> 15</p><p><b>Summary:</b> As LLM-based agents are increasingly deployed in real-life scenarios, existing benchmarks fail to capture their inherent complexity of handling extensive information, leveraging diverse resources, and managing dynamic user interactions. To address this gap, we introduce VitaBench, a challenging benchmark that evaluates agents on versatile interactive tasks grounded in real-world settings. Drawing from daily applications in food delivery, in-store consumption, and online travel services, VitaBench presents agents with the most complex life-serving simulation environment to date, comprising 66 tools. Through a framework that eliminates domain-specific policies, we enable flexible composition of these scenarios and tools, yielding 100 cross-scenario tasks (main results) and 300 single-scenario tasks. Each task is derived from multiple real user requests and requires agents to reason across temporal and spatial dimensions, utilize complex tool sets, proactively clarify ambiguous instructions, and track shifting user intent throughout multi-turn conversations. Moreover, we propose a rubric-based sliding window evaluator, enabling robust assessment of diverse solution pathways in complex environments and stochastic interactions. Our comprehensive evaluation reveals that even the most advanced models achieve only 30% success rate on cross-scenario tasks, and less than 50% success rate on others. Overall, we believe VitaBench will serve as a valuable resource for advancing the development of AI agents in practical real-world applications. The code, dataset, and leaderboard are available at https://vitabench.github.io/</p> + https://arxiv.org/abs/2509.26490 + Tue, 30 Sep 2025 16:33:49 +0000 + + + OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always! + https://arxiv.org/abs/2509.26495 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26495.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Jingdi Lei, Varun Gumma, Rishabh Bhardwaj, Seok Min Lim, Chuan Li, Amir Zadeh, Soujanya Poria</p><p><b>Upvotes:</b> 9</p><p><b>Summary:</b> Large Language Model (LLM) safety is one of the most pressing challenges for enabling wide-scale deployment. While most studies and global discussions focus on generic harms, such as models assisting users in harming themselves or others, enterprises face a more fundamental concern: whether LLM-based agents are safe for their intended use case. To address this, we introduce operational safety, defined as an LLM's ability to appropriately accept or refuse user queries when tasked with a specific purpose. We further propose OffTopicEval, an evaluation suite and benchmark for measuring operational safety both in general and within specific agentic use cases. Our evaluations on six model families comprising 20 open-weight LLMs reveal that while performance varies across models, all of them remain highly operationally unsafe. Even the strongest models -- Qwen-3 (235B) with 77.77\% and Mistral (24B) with 79.96\% -- fall far short of reliable operational safety, while GPT models plateau in the 62--73\% range, Phi achieves only mid-level scores (48--70\%), and Gemma and Llama-3 collapse to 39.53\% and 23.84\%, respectively. While operational safety is a core model alignment issue, to suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23\%, while P-ground delivers even larger boosts, raising Llama-3.3 (70B) by 41\% and Qwen-3 (30B) by 27\%. These results highlight both the urgent need for operational safety interventions and the promise of prompt-based steering as a first step toward more reliable LLM-based agents.</p> + https://arxiv.org/abs/2509.26495 + Tue, 30 Sep 2025 16:39:17 +0000 + + + The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain + https://arxiv.org/abs/2509.26507 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26507.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Adrian Kosowski, Przemysław Uznański, Jan Chorowski, Zuzanna Stamirowska, Michał Bartoszkiewicz</p><p><b>Upvotes:</b> 105</p><p><b>Summary:</b> The relationship between computing systems and the brain has served as motivation for pioneering theoreticians since John von Neumann and Alan Turing. Uniform, scale-free biological networks, such as the brain, have powerful properties, including generalizing over time, which is the main barrier for Machine Learning on the path to Universal Reasoning Models. We introduce `Dragon Hatchling' (BDH), a new Large Language Model architecture based on a scale-free biologically inspired network of \n locally-interacting neuron particles. BDH couples strong theoretical foundations and inherent interpretability without sacrificing Transformer-like performance. BDH is a practical, performant state-of-the-art attention-based state space sequence learning architecture. In addition to being a graph model, BDH admits a GPU-friendly formulation. It exhibits Transformer-like scaling laws: empirically BDH rivals GPT2 performance on language and translation tasks, at the same number of parameters (10M to 1B), for the same training data. BDH can be represented as a brain model. The working memory of BDH during inference entirely relies on synaptic plasticity with Hebbian learning using spiking neurons. We confirm empirically that specific, individual synapses strengthen connection whenever BDH hears or reasons about a specific concept while processing language inputs. The neuron interaction network of BDH is a graph of high modularity with heavy-tailed degree distribution. The BDH model is biologically plausible, explaining one possible mechanism which human neurons could use to achieve speech. BDH is designed for interpretability. Activation vectors of BDH are sparse and positive. We demonstrate monosemanticity in BDH on language tasks. Interpretability of state, which goes beyond interpretability of neurons and model parameters, is an inherent feature of the BDH architecture.</p> + https://arxiv.org/abs/2509.26507 + Tue, 30 Sep 2025 16:49:01 +0000 + + + OceanGym: A Benchmark Environment for Underwater Embodied Agents + https://arxiv.org/abs/2509.26536 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26536.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yida Xue, Mingjun Mao, Xiangyuan Ru, Yuqi Zhu, Baochang Ren, Shuofei Qiao, Mengru Wang, Shumin Deng, Xinyu An, Ningyu Zhang, Ying Chen, Huajun Chen</p><p><b>Upvotes:</b> 29</p><p><b>Summary:</b> We introduce OceanGym, the first comprehensive benchmark for ocean underwater embodied agents, designed to advance AI in one of the most demanding real-world environments. Unlike terrestrial or aerial domains, underwater settings present extreme perceptual and decision-making challenges, including low visibility, dynamic ocean currents, making effective agent deployment exceptionally difficult. OceanGym encompasses eight realistic task domains and a unified agent framework driven by Multi-modal Large Language Models (MLLMs), which integrates perception, memory, and sequential decision-making. Agents are required to comprehend optical and sonar data, autonomously explore complex environments, and accomplish long-horizon objectives under these harsh conditions. Extensive experiments reveal substantial gaps between state-of-the-art MLLM-driven agents and human experts, highlighting the persistent difficulty of perception, planning, and adaptability in ocean underwater environments. By providing a high-fidelity, rigorously designed platform, OceanGym establishes a testbed for developing robust embodied AI and transferring these capabilities to real-world autonomous ocean underwater vehicles, marking a decisive step toward intelligent agents capable of operating in one of Earth's last unexplored frontiers. The code and data are available at https://github.com/OceanGPT/OceanGym.</p> + https://arxiv.org/abs/2509.26536 + Tue, 30 Sep 2025 17:09:32 +0000 + + + Ferret-UI Lite: Lessons from Building Small On-Device GUI Agents + https://arxiv.org/abs/2509.26539 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26539.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Zhen Yang, Zi-Yi Dou, Di Feng, Forrest Huang, Anh Nguyen, Keen You, Omar Attia, Yuhao Yang, Michael Feng, Haotian Zhang, Ram Ramrakhya, Chao Jia, Jeffrey Nichols, Alexander Toshev, Yinfei Yang, Zhe Gan</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Developing autonomous agents that effectively interact with Graphic User Interfaces (GUIs) remains a challenging open problem, especially for small on-device models. In this paper, we present Ferret-UI Lite, a compact, end-to-end GUI agent that operates across diverse platforms, including mobile, web, and desktop. Utilizing techniques optimized for developing small models, we build our 3B Ferret-UI Lite agent through curating a diverse GUI data mixture from real and synthetic sources, strengthening inference-time performance through chain-of-thought reasoning and visual tool-use, and reinforcement learning with designed rewards. Ferret-UI Lite achieves competitive performance with other small-scale GUI agents. In GUI grounding, Ferret-UI Lite attains scores of 91.6%, 53.3%, and 61.2% on the ScreenSpot-V2, ScreenSpot-Pro, and OSWorld-G benchmarks, respectively. For GUI navigation, Ferret-UI Lite achieves success rates of 28.0% on AndroidWorld and 19.8% on OSWorld. We share our methods and lessons learned from developing compact, on-device GUI agents.</p> + https://arxiv.org/abs/2509.26539 + Tue, 30 Sep 2025 17:13:56 +0000 + + + Voice Evaluation of Reasoning Ability: Diagnosing the Modality-Induced Performance Gap + https://arxiv.org/abs/2509.26542 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26542.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yueqian Lin, Zhengmian Hu, Qinsi Wang, Yudong Liu, Hengfan Zhang, Jayakumar Subramanian, Nikos Vlassis, Hai Helen Li, Yiran Chen</p><p><b>Upvotes:</b> 8</p><p><b>Summary:</b> We present Voice Evaluation of Reasoning Ability (VERA), a benchmark for evaluating reasoning ability in voice-interactive systems under real-time conversational constraints. VERA comprises 2,931 voice-native episodes derived from established text benchmarks and organized into five tracks (Math, Web, Science, Long-Context, Factual). Each item is adapted for speech interaction while preserving reasoning difficulty. VERA enables direct text-voice comparison within model families and supports analysis of how architectural choices affect reliability. We assess 12 contemporary voice systems alongside strong text baselines and observe large, consistent modality gaps: on competition mathematics a leading text model attains 74.8% accuracy while its voice counterpart reaches 6.1%; macro-averaged across tracks the best text models achieve 54.0% versus 11.3% for voice. Latency-accuracy analyses reveal a low-latency plateau, where fast voice systems cluster around ~10% accuracy, while approaching text performance requires sacrificing real-time interaction. Diagnostic experiments indicate that common mitigations are insufficient. Increasing "thinking time" yields negligible gains; a decoupled cascade that separates reasoning from narration improves accuracy but still falls well short of text and introduces characteristic grounding/consistency errors. Failure analyses further show distinct error signatures across native streaming, end-to-end, and cascade designs. VERA provides a reproducible testbed and targeted diagnostics for architectures that decouple thinking from speaking, offering a principled way to measure progress toward real-time voice assistants that are both fluent and reliably reasoned.</p> + https://arxiv.org/abs/2509.26542 + Tue, 30 Sep 2025 17:17:09 +0000 + + + Stable Cinemetrics : Structured Taxonomy and Evaluation for Professional Video Generation + https://arxiv.org/abs/2509.26555 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26555.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Agneet Chatterjee, Rahim Entezari, Maksym Zhuravinskyi, Maksim Lapin, Reshinth Adithyan, Amit Raj, Chitta Baral, Yezhou Yang, Varun Jampani</p><p><b>Upvotes:</b> 0</p><p><b>Summary:</b> Recent advances in video generation have enabled high-fidelity video synthesis from user provided prompts. However, existing models and benchmarks fail to capture the complexity and requirements of professional video generation. Towards that goal, we introduce Stable Cinemetrics, a structured evaluation framework that formalizes filmmaking controls into four disentangled, hierarchical taxonomies: Setup, Event, Lighting, and Camera. Together, these taxonomies define 76 fine-grained control nodes grounded in industry practices. Using these taxonomies, we construct a benchmark of prompts aligned with professional use cases and develop an automated pipeline for prompt categorization and question generation, enabling independent evaluation of each control dimension. We conduct a large-scale human study spanning 10+ models and 20K videos, annotated by a pool of 80+ film professionals. Our analysis, both coarse and fine-grained reveal that even the strongest current models exhibit significant gaps, particularly in Events and Camera-related controls. To enable scalable evaluation, we train an automatic evaluator, a vision-language model aligned with expert annotations that outperforms existing zero-shot baselines. SCINE is the first approach to situate professional video generation within the landscape of video generative models, introducing taxonomies centered around cinematic controls and supporting them with structured evaluation pipelines and detailed analyses to guide future research.</p> + https://arxiv.org/abs/2509.26555 + Tue, 30 Sep 2025 17:22:18 +0000 + + + Probing the Critical Point (CritPt) of AI Reasoning: a Frontier Physics Research Benchmark + https://arxiv.org/abs/2509.26574 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26574.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Minhui Zhu, Minyang Tian, Xiaocheng Yang, Tianci Zhou, Penghao Zhu, Eli Chertkov, Shengyan Liu, Yufeng Du, Lifan Yuan, Ziming Ji, Indranil Das, Junyi Cao, Yufeng Du, Jinchen He, Yifan Su, Jiabin Yu, Yikun Jiang, Yujie Zhang, Chang Liu, Ze-Min Huang, Weizhen Jia, Xinan Chen, Peixue Wu, Yunkai Wang, Juntai Zhou, Yong Zhao, Farshid Jafarpour, Jessie Shelton, Aaron Young, John Bartolotta, Wenchao Xu, Yue Sun, Anjun Chu, Victor Colussi, Chris Akers, Nathan Brooks, Wenbo Fu, Christopher Wilson, Jinchao Zhao, Marvin Qi, Anqi Mu, Yubo Yang, Allen Zang, Yang Lyu, Peizhi Mai, Xuefei Guo, Luyu Gao, Ze Yang, Chi Xue, Dmytro Bandak, Yaïr Hein, Yonatan Kahn, Kevin Zhou, John Drew Wilson Jarrod T. Reilly, Di Luo, Daniel Inafuku, Hao Tong, Liang Yang, Ruixing Zhang, Xueying Wang, Ofir Press, Nicolas Chia, Eliu Huerta, Hao Peng</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> While large language models (LLMs) with reasoning capabilities are progressing rapidly on high-school math competitions and coding, can they reason effectively through complex, open-ended challenges found in frontier physics research? And crucially, what kinds of reasoning tasks do physicists want LLMs to assist with? To address these questions, we present the CritPt (Complex Research using Integrated Thinking - Physics Test, pronounced "critical point"), the first benchmark designed to test LLMs on unpublished, research-level reasoning tasks that broadly covers modern physics research areas, including condensed matter, quantum physics, atomic, molecular & optical physics, astrophysics, high energy physics, mathematical physics, statistical physics, nuclear physics, nonlinear dynamics, fluid dynamics and biophysics. CritPt consists of 71 composite research challenges designed to simulate full-scale research projects at the entry level, which are also decomposed to 190 simpler checkpoint tasks for more fine-grained insights. All problems are newly created by 50+ active physics researchers based on their own research. Every problem is hand-curated to admit a guess-resistant and machine-verifiable answer and is evaluated by an automated grading pipeline heavily customized for advanced physics-specific output formats. We find that while current state-of-the-art LLMs show early promise on isolated checkpoints, they remain far from being able to reliably solve full research-scale challenges: the best average accuracy among base models is only 4.0% , achieved by GPT-5 (high), moderately rising to around 10% when equipped with coding tools. Through the realistic yet standardized evaluation offered by CritPt, we highlight a large disconnect between current model capabilities and realistic physics research demands, offering a foundation to guide the development of scientifically grounded AI tools.</p> + https://arxiv.org/abs/2509.26574 + Tue, 30 Sep 2025 17:34:03 +0000 + + + DeepScientist: Advancing Frontier-Pushing Scientific Findings Progressively + https://arxiv.org/abs/2509.26603 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26603.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Yixuan Weng, Minjun Zhu, Qiujie Xie, Qiyao Sun, Zhen Lin, Sifan Liu, Yue Zhang</p><p><b>Upvotes:</b> 12</p><p><b>Summary:</b> While previous AI Scientist systems can generate novel findings, they often lack the focus to produce scientifically valuable contributions that address pressing human-defined challenges. We introduce DeepScientist, a system designed to overcome this by conducting goal-oriented, fully autonomous scientific discovery over month-long timelines. It formalizes discovery as a Bayesian Optimization problem, operationalized through a hierarchical evaluation process consisting of "hypothesize, verify, and analyze". Leveraging a cumulative Findings Memory, this loop intelligently balances the exploration of novel hypotheses with exploitation, selectively promoting the most promising findings to higher-fidelity levels of validation. Consuming over 20,000 GPU hours, the system generated about 5,000 unique scientific ideas and experimentally validated approximately 1100 of them, ultimately surpassing human-designed state-of-the-art (SOTA) methods on three frontier AI tasks by 183.7\%, 1.9\%, and 7.9\%. This work provides the first large-scale evidence of an AI achieving discoveries that progressively surpass human SOTA on scientific tasks, producing valuable findings that genuinely push the frontier of scientific discovery. To facilitate further research into this process, we will open-source all experimental logs and system code at https://github.com/ResearAI/DeepScientist/.</p> + https://arxiv.org/abs/2509.26603 + Tue, 30 Sep 2025 17:49:32 +0000 + + + Video Object Segmentation-Aware Audio Generation + https://arxiv.org/abs/2509.26604 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26604.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Ilpo Viertola, Vladimir Iashin, Esa Rahtu</p><p><b>Upvotes:</b> 1</p><p><b>Summary:</b> Existing multimodal audio generation models often lack precise user control, which limits their applicability in professional Foley workflows. In particular, these models focus on the entire video and do not provide precise methods for prioritizing a specific object within a scene, generating unnecessary background sounds, or focusing on the wrong objects. To address this gap, we introduce the novel task of video object segmentation-aware audio generation, which explicitly conditions sound synthesis on object-level segmentation maps. We present SAGANet, a new multimodal generative model that enables controllable audio generation by leveraging visual segmentation masks along with video and textual cues. Our model provides users with fine-grained and visually localized control over audio generation. To support this task and further research on segmentation-aware Foley, we propose Segmented Music Solos, a benchmark dataset of musical instrument performance videos with segmentation information. Our method demonstrates substantial improvements over current state-of-the-art methods and sets a new standard for controllable, high-fidelity Foley synthesis. Code, samples, and Segmented Music Solos are available at https://saganet.notion.site</p> + https://arxiv.org/abs/2509.26604 + Tue, 30 Sep 2025 17:49:41 +0000 + + + DA^2: Depth Anything in Any Direction + https://arxiv.org/abs/2509.26618 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26618.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Haodong Li, Wangguangdong Zheng, Jing He, Yuhao Liu, Xin Lin, Xin Yang, Ying-Cong Chen, Chunchao Guo</p><p><b>Upvotes:</b> 11</p><p><b>Summary:</b> Panorama has a full FoV (360^circtimes180^circ), offering a more complete visual description than perspective images. Thanks to this characteristic, panoramic depth estimation is gaining increasing traction in 3D vision. However, due to the scarcity of panoramic data, previous methods are often restricted to in-domain settings, leading to poor zero-shot generalization. Furthermore, due to the spherical distortions inherent in panoramas, many approaches rely on perspective splitting (e.g., cubemaps), which leads to suboptimal efficiency. To address these challenges, we propose DA^{2}: Depth Anything in Any Direction, an accurate, zero-shot generalizable, and fully end-to-end panoramic depth estimator. Specifically, for scaling up panoramic data, we introduce a data curation engine for generating high-quality panoramic depth data from perspective, and create sim543K panoramic RGB-depth pairs, bringing the total to sim607K. To further mitigate the spherical distortions, we present SphereViT, which explicitly leverages spherical coordinates to enforce the spherical geometric consistency in panoramic image features, yielding improved performance. A comprehensive benchmark on multiple datasets clearly demonstrates DA^{2}'s SoTA performance, with an average 38% improvement on AbsRel over the strongest zero-shot baseline. Surprisingly, DA^{2} even outperforms prior in-domain methods, highlighting its superior zero-shot generalization. Moreover, as an end-to-end solution, DA^{2} exhibits much higher efficiency over fusion-based approaches. Both the code and the curated panoramic data will be released. Project page: https://depth-any-in-any-dir.github.io/.</p> + https://arxiv.org/abs/2509.26618 + Tue, 30 Sep 2025 17:55:37 +0000 + + + Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training + https://arxiv.org/abs/2509.26625 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26625.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Junlin Han, Shengbang Tong, David Fan, Yufan Ren, Koustuv Sinha, Philip Torr, Filippos Kokkinos</p><p><b>Upvotes:</b> 34</p><p><b>Summary:</b> Large Language Models (LLMs), despite being trained on text alone, surprisingly develop rich visual priors. These priors allow latent visual capabilities to be unlocked for vision tasks with a relatively small amount of multimodal data, and in some cases, to perform visual tasks without ever having seen an image. Through systematic analysis, we reveal that visual priors-the implicit, emergent knowledge about the visual world acquired during language pre-training-are composed of separable perception and reasoning priors with unique scaling trends and origins. We show that an LLM's latent visual reasoning ability is predominantly developed by pre-training on reasoning-centric data (e.g., code, math, academia) and scales progressively. This reasoning prior acquired from language pre-training is transferable and universally applicable to visual reasoning. In contrast, a perception prior emerges more diffusely from broad corpora, and perception ability is more sensitive to the vision encoder and visual instruction tuning data. In parallel, text describing the visual world proves crucial, though its performance impact saturates rapidly. Leveraging these insights, we propose a data-centric recipe for pre-training vision-aware LLMs and verify it in 1T token scale pre-training. Our findings are grounded in over 100 controlled experiments consuming 500,000 GPU-hours, spanning the full MLLM construction pipeline-from LLM pre-training to visual alignment and supervised multimodal fine-tuning-across five model scales, a wide range of data categories and mixtures, and multiple adaptation setups. Along with our main findings, we propose and investigate several hypotheses, and introduce the Multi-Level Existence Bench (MLE-Bench). Together, this work provides a new way of deliberately cultivating visual priors from language pre-training, paving the way for the next generation of multimodal LLMs.</p> + https://arxiv.org/abs/2509.26625 + Tue, 30 Sep 2025 17:57:44 +0000 + + + Attention as a Compass: Efficient Exploration for Process-Supervised RL in Reasoning Models + https://arxiv.org/abs/2509.26628 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26628.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Runze Liu, Jiakang Wang, Yuling Shi, Zhihui Xie, Chenxin An, Kaiyan Zhang, Jian Zhao, Xiaodong Gu, Lei Lin, Wenping Hu, Xiu Li, Fuzheng Zhang, Guorui Zhou, Kun Gai</p><p><b>Upvotes:</b> 7</p><p><b>Summary:</b> Reinforcement Learning (RL) has shown remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). Process-Supervised RL (PSRL) has emerged as a more effective paradigm compared to outcome-based RL. However, existing PSRL approaches suffer from limited exploration efficiency, both in terms of branching positions and sampling. In this paper, we introduce a novel PSRL framework (AttnRL), which enables efficient exploration for reasoning models. Motivated by preliminary observations that steps exhibiting high attention scores correlate with reasoning behaviors, we propose to branch from positions with high values. Furthermore, we develop an adaptive sampling strategy that accounts for problem difficulty and historical batch size, ensuring that the whole training batch maintains non-zero advantage values. To further improve sampling efficiency, we design a one-step off-policy training pipeline for PSRL. Extensive experiments on multiple challenging mathematical reasoning benchmarks demonstrate that our method consistently outperforms prior approaches in terms of performance and sampling and training efficiency.</p> + https://arxiv.org/abs/2509.26628 + Tue, 30 Sep 2025 17:58:34 +0000 + + + TTT3R: 3D Reconstruction as Test-Time Training + https://arxiv.org/abs/2509.26645 + <p><img src="https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2509.26645.png" alt="Paper thumbnail" style="max-width: 300px; height: auto;" /></p><p><b>Authors:</b> Xingyu Chen, Yue Chen, Yuliang Xiu, Andreas Geiger, Anpei Chen</p><p><b>Upvotes:</b> 4</p><p><b>Summary:</b> Modern Recurrent Neural Networks have become a competitive architecture for 3D reconstruction due to their linear-time complexity. However, their performance degrades significantly when applied beyond the training context length, revealing limited length generalization. In this work, we revisit the 3D reconstruction foundation models from a Test-Time Training perspective, framing their designs as an online learning problem. Building on this perspective, we leverage the alignment confidence between the memory state and incoming observations to derive a closed-form learning rate for memory updates, to balance between retaining historical information and adapting to new observations. This training-free intervention, termed TTT3R, substantially improves length generalization, achieving a 2times improvement in global pose estimation over baselines, while operating at 20 FPS with just 6 GB of GPU memory to process thousands of images. Code available in https://rover-xingyu.github.io/TTT3R</p> + https://arxiv.org/abs/2509.26645 + Tue, 30 Sep 2025 17:59:51 +0000