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。
50 篇论文 · 按点赞排序
Yi Chen, Yuying Ge, Rui Wang +4 authors
A benchmark (SEED-Bench-R1) evaluates reinforcement learning versus supervised fine-tuning for multimodal large language models in video understanding, highlighting RL's data efficiency and superior performance but identifying limitations in logical coherence and visual cue processing.
Alexander Martin, Reno Kriz, William Gantt Walden +5 authors
WikiVideo proposes Collaborative Article Generation (CAG) to enhance high-level event summarization from videos by integrating r1-style reasoning and VideoLLM for better inferences compared to state-of-the-art VideoLLMs.
Giulio Starace, Oliver Jaffe, Dane Sherburn +10 authors
PaperBench evaluates AI agents' ability to replicate state-of-the-art AI research by decomposing replication tasks into graded sub-tasks, using both LLM-based and human judges to assess performance.
Jiakai Tang, Sunhao Dai, Teng Shi +5 authors
ReaRec enhances sequential recommendation by using inference-time multi-step reasoning, improving model performance on long-tail items and user preference evolution.
Anjiang Wei, Tarun Suresh, Jiannan Cao +6 authors
CodeARC introduces an interactive evaluation framework to assess the ability of large language model agents in inductive program synthesis using differential testing and iterative refinement.
David Fan, Shengbang Tong, Jiachen Zhu +8 authors
Visual self-supervised learning matches language-supervised visual pretraining performance on VQA and vision benchmarks when both are trained on the same dataset and scaled appropriately.
Yan Ma, Steffi Chern, Xuyang Shen +2 authors
A transparent reinforcement learning framework and standardized evaluation are introduced for vision-language models, showing RL's superiority in generalization over supervised fine-tuning.
Team Cohere, Aakanksha, Arash Ahmadian +223 authors
Command A, a multilingual large language model, uses decentralized training with self-refinement and model merging to achieve efficient and top-performing Retrieval Augmented Generation for enterprise use.
Gallil Maimon, Michael Hassid, Amit Roth +1 authors
Interleaved Speech Language Models scale more efficiently than textless models, achieving comparable performance with less compute and data.
Tian-Xing Xu, Xiangjun Gao, Wenbo Hu +3 authors
GeometryCrafter uses a point map VAE and video diffusion model to estimate high-fidelity, temporally coherent depth maps from open-world videos, enhancing 3D reconstruction and camera parameter estimation.
Zhiyuan Liu, Yuting Zhang, Feng Liu +3 authors
OThink-MR1, an advanced MLLM using GRPO-D, enhances reinforcement learning performance and demonstrates superior cross-task generalization compared to supervised fine-tuning.
Zhaojian Yu, Yinghao Wu, Yilun Zhao +2 authors
A novel Shifted Thinking Window method trains LLMs on code-related reasoning trajectories to reduce excessive thinking tokens while maintaining performance and efficient test-time scaling.
Saaket Agashe, Kyle Wong, Vincent Tu +3 authors
Agent S2, a compositional framework using Mixture-of-Grounding and Proactive Hierarchical Planning, achieves state-of-the-art performance in computer use automation across various benchmarks and operating systems.
Zhonghan Zhao, Wenwei Zhang, Haian Huang +4 authors
The paper presents RIG, an end-to-end agent policy that integrates reasoning and imagination, significantly improving sample efficiency and generalization in complex environments through joint reasoning and action outcome prediction.
Qianhao Yuan, Qingyu Zhang, Yanjiang Liu +6 authors
ShortV identifies and freezes ineffective layers in MLLMs to reduce computational costs without significant performance loss.
Zhanke Zhou, Zhaocheng Zhu, Xuan Li +5 authors
A visualization tool named Landscape of Thoughts helps inspect and analyze reasoning paths of large language models on multi-choice datasets, identifying model strengths, correct answers, and reasoning inconsistencies.
Tommie Kerssies, Niccolò Cavagnero, Alexander Hermans +5 authors
The Encoder-only Mask Transformer (EoMT) achieves state-of-the-art image segmentation accuracy by learning task-specific components through large-scale pre-training, while maintaining superior prediction speed compared to models with additional architectural complexity.
Taewon Yun, Jihwan Oh, Hyangsuk Min +4 authors
Summarization refinement faces challenges when extending to multi-dimension. In this paper, we introduce ReFeed, a powerful summarization refinement pipeline that enhances multiple dimensions through reflective reasoning on feedback. To achieve this, we release SumFeed-CoT, a large-scale Long-CoT-based dataset optimized for training a lightweight model with reflective reasoning. Our experiments reveal how the number of dimensions, feedback exposure, and reasoning policy influence refinement performance, highlighting reflective reasoning and simultaneously addressing multiple feedback is crucial to mitigate trade-off between dimensions. Furthermore, ReFeed is robust to noisy feedback and feedback order. Lastly, our finding emphasizes that creating data with a proper goal and guideline constitutes a fundamental pillar of effective reasoning. The dataset and model will be released.
Kai Yan, Yufei Xu, Zhengyin Du +4 authors
Existing large language models exhibit significant recitation behavior when conditions are subtly altered, suggesting they may not possess human-like intelligence.
Yi Su, Dian Yu, Linfeng Song +5 authors
Reinforcement learning with verifiable rewards extended to diverse domains using model-based soft scoring outperforms existing LLMs in free-form answer settings with reliable reward signals.
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