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。
Keke Lian, Bin Wang, Lei Zhang +18 authors
A.S.E is a repository-level benchmark for evaluating the security of AI-generated code, highlighting challenges in secure coding and the limitations of LLMs in real-world scenarios.
50 篇论文 · 按点赞排序
Guibin Zhang, Hejia Geng, Xiaohang Yu +22 authors
Agentic reinforcement learning transforms large language models into autonomous decision-making agents by leveraging temporally extended POMDPs, enhancing capabilities like planning and reasoning through reinforcement learning.
Yang Wang, Chenghao Xiao, Chia-Yi Hsiao +4 authors
LLMs struggle with understanding the nuanced, context-dependent meanings of Drivelological text, which appears nonsensical but contains deeper semantic layers.
Ming Hu, Chenglong Ma, Wei Li +100 authors
Sci-LLMs are evolving through a co-development with scientific data, addressing unique challenges like multimodal and domain-specific information, and are moving towards autonomous, closed-loop systems in scientific research.
Haoming Wang, Haoyang Zou, Huatong Song +103 authors
UI-TARS-2, a native GUI-centered agent model, addresses challenges in data scalability, multi-turn reinforcement learning, and environment stability, achieving significant improvements over its predecessor and strong baselines across various benchmarks.
Jie Jiang, Qi Yang, Bolin Ni +3 authors
R-4B, an auto-thinking multimodal large language model, uses bi-mode annealing and Bi-mode Policy Optimization to adaptively decide on problem-solving strategies, achieving state-of-the-art performance with lower computational cost.
JiYuan Wang, Chunyu Lin, Lei Sun +6 authors
FE2E, a framework using a Diffusion Transformer for dense geometry prediction, outperforms generative models in zero-shot monocular depth and normal estimation with improved performance and efficiency.
Xiyao Wang, Chunyuan Li, Jianwei Yang +4 authors
Reinforcement learning on preference-labeled critic datasets enhances a generative model's performance, creating a unified multimodal system that excels in both evaluation and generation.
Zhenghai Xue, Longtao Zheng, Qian Liu +4 authors
SimpleTIR stabilizes multi-turn Tool-Integrated Reasoning training by filtering out void turns, achieving state-of-the-art performance on math reasoning benchmarks.
Dongfu Jiang, Yi Lu, Zhuofeng Li +9 authors
VerlTool is a unified and modular framework for Agentic Reinforcement Learning with Tool use, addressing inefficiencies in existing approaches and providing competitive performance across multiple domains.
Delin Qu, Haoming Song, Qizhi Chen +12 authors
EO-Robotics, comprising EO-1 model and EO-Data1.5M dataset, advances multimodal embodied reasoning and robot control through interleaved vision-text-action pre-training.
Xingtai Lv, Yuxin Zuo, Youbang Sun +9 authors
A unified policy gradient estimator and Hybrid Post-Training algorithm effectively combine online and offline data for post-training language models, improving performance across various benchmarks.
Xiaochuan Li, Guoguang Du, Runze Zhang +11 authors
Using video data to provide commonsense priors enhances 3D asset generation, enabling spatial consistency and semantic plausibility in 3D content creation.
Ziyi Xia, Kun Luo, Hongjin Qian +1 authors
InfoSeek is a scalable framework for generating complex Deep Research tasks by synthesizing hierarchical constraint satisfaction problems, enabling models to outperform larger baselines on challenging benchmarks.
Mohammad Zbeeb, Hasan Abed Al Kader Hammoud, Bernard Ghanem
Reasoning capabilities from reinforcement learning can be extracted as a task vector and transferred to other models to improve performance on diverse benchmarks.
Haiyuan Wan, Chen Yang, Junchi Yu +10 authors
DeepResearch Arena, a benchmark using academic seminar transcripts, provides high-quality research tasks to evaluate deep research agents across multiple disciplines.
Qinyan Zhang, Xinping Lei, Ruijie Miao +18 authors
Inverse IFEval evaluates Large Language Models' ability to override training biases and follow unconventional instructions, highlighting the need for adaptability in diverse contexts.
Hao Lu, Jiahao Wang, Yaolun Zhang +5 authors
A benchmark for long-video hallucination identifies and investigates Semantic Aggregation Hallucination (SAH), showing its prevalence in complex and rapidly changing semantic contexts, and proposes strategies to mitigate it.
Huang Fang, Mengxi Zhang, Heng Dong +6 authors
Robix, a unified vision-language model, integrates robot reasoning, task planning, and natural language interaction, demonstrating superior performance in interactive task execution through chain-of-thought reasoning and a three-stage training strategy.
Yuan Liu, Zhongyin Zhao, Le Tian +8 authors
A framework for constructing high-quality document extraction datasets and models through synthetic data generation and iterative self-improvement outperforms existing models.
Baichuan-M2 Team, Chengfeng Dou, Chong Liu +31 authors
A dynamic verification framework using reinforcement learning and a novel algorithm improves the performance of a large medical language model in real-world clinical decision-making.
Rishiraj Acharya
Gated Associative Memory (GAM) network offers a linear complexity alternative to the Transformer architecture, improving training speed and achieving competitive validation perplexity.
Biao Yang, Bin Wen, Boyang Ding +57 authors
Keye-VL-1.5 enhances video understanding through a Slow-Fast encoding strategy, progressive pre-training, and post-training reasoning improvements, outperforming existing models on video tasks while maintaining general multimodal performance.
Wenfeng Feng, Penghong Zhao, Guochao Jiang +3 authors
PVPO, an enhanced critic-free reinforcement learning method, uses a reference model for pre-sampling and advantage estimation, improving efficiency and performance across various tasks and model scales.
Yanqing Liu, Xianhang Li, Letian Zhang +4 authors
OpenVision 2 simplifies the original architecture by removing the text encoder and contrastive loss, achieving competitive performance with reduced training time and memory usage, and enabling scaling to over 1 billion parameters.
Zidong Wang, Yiyuan Zhang, Xiaoyu Yue +4 authors
A novel generative paradigm, Transition Models (TiM), addresses the trade-off between computational cost and output quality in generative modeling by using a continuous-time dynamics equation.
Or Shachar, Uri Katz, Yoav Goldberg +1 authors
NER Retriever uses internal representations from large language models to perform zero-shot named entity retrieval by embedding entity mentions and type descriptions into a shared semantic space, outperforming lexical and dense sentence-level retrieval methods.
Yi Liao, Yu Gu, Yuan Sui +5 authors
Think in Games (TiG) framework enables large language models to develop procedural knowledge through interactive game environments, achieving competitive performance with reduced data and computational demands while providing transparent explanations.
Jie Zhang, Changzai Pan, Kaiwen Wei +12 authors
A bilingual benchmark named T2R-bench is proposed to evaluate the performance of large language models in generating reports from tables, highlighting the need for improvement in this task.
Xingyue Huang, Rishabh, Gregor Franke +43 authors
The Loong Project introduces a framework for generating and verifying synthetic data to improve reasoning capabilities in Large Language Models through Reinforcement Learning with Verifiable Reward.
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