Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
686 篇论文 · 按点赞排序
Dian Zheng, Harry Lee, Manyuan Zhang +4 authors
InterleaveThinker enables interleaved generation capabilities for image generators through a multi-agent pipeline with planner and critic agents, achieving performance comparable to state-of-the-art models while enhancing reasoning benchmarks.
Bo-Wen Zhang, Junwei He, Wen Wang +5 authors
CoRT improves rubric-based reinforcement learning by using counterfactual replay to assign token-level credit weights within responses, boosting performance without auxiliary scorers.
Jian Zhang, Yu He, Zhiyuan Wang +5 authors
Scientific reasoning relies not only on logical inference but also on activating prior knowledge and experiential structures. Memory can efficiently reuse knowledge and enhance reasoning consistency and stability. However, existing benchmarks mainly evaluate final answers or step-by-step coherence, overlooking the memory-driven mechanisms that underlie human reasoning, which involves activating anchors and attractors, then integrating them into multi-step inference. To address this gap, we propose A^3-Bench~ https://a3-bench.github.io, a benchmark designed to evaluate scientific reasoning through dual-scale memory-driven activation, grounded in Anchor and Attractor Activation. First, we annotate 2,198 science reasoning problems across domains using the SAPM process(subject, anchor & attractor, problem, and memory developing). Second, we introduce a dual-scale memory evaluation framework utilizing anchors and attractors, along with the AAUI(Anchor--Attractor Utilization Index) metric to measure memory activation rates. Finally, through experiments with various base models and paradigms, we validate A^3-Bench and analyze how memory activation impacts reasoning performance, providing insights into memory-driven scientific reasoning.
Guixu Lin, Zheng-Hui Huang, Siqi Yang +3 authors
AlayaRenderer-Flash accelerates a generative world renderer to real-time speeds via few-step autoregressive streaming and distilled codecs while preserving structured scene dynamics.
Chunchao Guo, Jinpeng Li, Yang Li +1 authors
WorldClaw is an agentic coarse-to-fine framework that generates large-scale editable 3D worlds from text by combining planning agents, semantic layouts, reusable assets, and render-based refinement.
Dianyi Wang, Ruihang Li, Feng Han +17 authors
DeepGen 1.0 is a lightweight 5B unified multimodal model that achieves superior performance in image generation and editing through advanced alignment techniques and progressive training strategies.
Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster +34 authors
AIRS-Bench presents a comprehensive benchmark suite for evaluating LLM agents across diverse scientific domains, demonstrating current limitations while providing open-source resources for advancement.
Chanyoung Kim, Minwoo Kim, Minseok Kang +2 authors
Vision-Language-Action models show significant performance drops when handling paraphrased instructions due to surface-level matching rather than semantic understanding, highlighting the need for better linguistic generalization metrics.
Lianghui Zhu, Yuxin Fang, Bencheng Liao +10 authors
Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .
Jiaqi Tang, Jianmin Chen, Youyang Zhai +6 authors
Robust-U1 enhances multimodal large language models' robustness against visual corruptions through self-recovery capabilities that improve both visual quality and reasoning performance.
Xinyu Wang, Chongbo Zhao, Fangneng Zhan +1 authors
A novel streaming video editing framework enables causal, frame-by-frame editing with stable long-horizon preservation and real-time responsiveness through a three-stage distillation pipeline and AR-oriented mask cache.
Qihao Wang, Ziming Cheng, Shuo Zhang +12 authors
MemGovern framework transforms unstructured GitHub data into structured experiential memory for autonomous software engineering agents, improving bug resolution rates through enhanced experience retrieval.
Zunhai Su, Hengyuan Zhang, Wei Wu +17 authors
Transformers face challenges from Attention Sink phenomenon where excessive attention focuses on uninformative tokens, impacting interpretability and performance, necessitating comprehensive research survey addressing fundamental usage, mechanistic understanding, and strategic mitigation approaches.
Jiajun Liang, Yucheng Liao, Yukang Cao +12 authors
AURORA-LM is a continuous-latent diffusion language model that preserves high-capacity decodable text representations and uses a block-causal diffusion transformer with flow matching to generate text.
Wei Zhou, Xiongwei Zhu, Zelin Xu +8 authors
A novel on-policy generative field distillation framework called DanceOPD is proposed to unify text-to-image generation, local editing, and global editing capabilities in flow-matching models through capability-specific routing and velocity-based training.
Sergii Kozyrev, Davyd Maiboroda
Fully quantizing hybrid LLMs—including recurrent Gated DeltaNet layers—to 4-bit NVFP4 preserves accuracy across long-context and reasoning benchmarks by localizing outliers and exploiting robust delta-rule dynamics.
Yunxin Li, Jinchao Li, Shibo Su +7 authors
KnowAct-GUIClaw introduces a Know-Route-Act-Reflect framework that enhances cross-platform GUI automation and recursive self-improvement through experience-driven memory and skill libraries.
Jinyuan Li, Chengsong Huang, Langlin Huang +4 authors
Training multimodal process reward models efficiently through balanced-information scoring that prioritizes label mixture and reliability while achieving full-data performance with only 10% of training data.
Jincheng Zhong, Weizhi Wang, Che Jiang +5 authors
EnterpriseClawBench presents a benchmark for enterprise agents based on real-world sessions with 852 reproducible tasks, emphasizing comprehensive evaluation metrics beyond single performance scores.
Xingyu Tan, Xiaoyang Wang, Qing Liu +4 authors
SkillZip compresses reusable procedural skills into contract-preserving, executable graph units to enable efficient retrieval and expansion under limited context budgets.
Yuhang Wang, Yuling Shi, Shaoqiu Zhang +6 authors
SWE-Pruner Pro uses an agent's internal representations to prune tool outputs directly, reducing token usage while maintaining or improving coding task performance.
Nikita Kachaev, Andrey Moskalenko, Matvey Skripkin +10 authors
Act2Answer protocol evaluates embodied vision-language-action models by having agents answer questions through physical actions, revealing knowledge retention and generalization patterns across different semantic categories.
Bingnan Li, Haozhe Wang, Haozhong Xiong +5 authors
On-policy diffusion distillation under classifier-free guidance suffers from branch-level error compensation, which is resolved by separately matching positive predictions and guidance directions.
Dianyi Wang, Chaofan Ma, Feng Han +8 authors
UniReason integrates text-to-image generation and image editing through a dual reasoning paradigm that enhances planning with world knowledge and uses editing for visual refinement, achieving superior performance on reasoning-intensive benchmarks.
Vishesh Tripathi, Abhay Kumar
Grouped Query Experts (GQE) improves Transformer efficiency by selectively activating query heads based on token content while maintaining key-value cache benefits of grouped-query attention.
Shiyang Feng, Runmin Ma, Xiangchao Yan +54 authors
InternAgent-1.5 is a unified system for autonomous scientific discovery that integrates computational modeling and experimental research through coordinated subsystems for generation, verification, and evolution.
Haoyu Zhao, Xingyue Zhao, Hangyu Li +6 authors
Digital teleoperation replaces physical robot interaction with generative world models to create diverse training data for robotics, enabling efficient zero-shot Sim2Real transfer and improved real-world performance.
Fukang Wen, Zhijie Wang, Ruilin Xu
OpenRath introduces a PyTorch-like programming model for multi-agent systems using Session as a central runtime abstraction that enables explicit fork, merge, and replay operations while recording comprehensive execution state.
NeoteAI Team, Fudan TEAI Team
A vision-tactile-language-action model pretrained on large-scale tactile data achieves fine-grained contact-rich manipulation and improves via advantage-conditioned offline reinforcement learning.
Lizhen Wang, Yongming Zhu, Zhipeng Ge +15 authors
FlowAct-R1 enables real-time interactive humanoid video generation with high-fidelity synthesis and low-latency responsiveness through MMDiT architecture and chunkwise diffusion forcing strategies.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号