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
Xinyuan Wang, Bowen Wang, Dunjie Lu +36 authors
OpenCUA is an open-source framework for vision-language models as computer-use agents, featuring an annotation infrastructure, a large-scale dataset, and a scalable pipeline that achieves state-of-the-art performance.
Zhitian Xie, Qintong Wu, Chengyue Yu +2 authors
A dynamic Multi-Agent System (MAS) with Execution and Guard Agents improves the reliability and effectiveness of intelligent agents using external tools, outperforming single-agent systems on the GAIA leaderboard.
Yushi Lan, Yihang Luo, Fangzhou Hong +7 authors
STream3R reformulates 3D reconstruction as a decoder-only Transformer problem, using causal attention to efficiently process image sequences and outperform existing methods in both static and dynamic scenes.
Xu Wang, Chenkai Xu, Yijie Jin +3 authors
Discrete diffusion forcing enhances diffusion large language models with block-wise autoregressive generation and inter-block parallel decoding, significantly improving inference speed while maintaining quality.
Cheng Qian, Zuxin Liu, Akshara Prabhakar +9 authors
UserBench evaluates LLM-based agents in multi-turn interactions with simulated users, revealing gaps in task completion and user alignment.
Weijia Wu, Chen Gao, Joya Chen +6 authors
This survey synthesizes recent advancements in visual reinforcement learning, covering policy optimization strategies, thematic pillars, and evaluation protocols, while highlighting open challenges.
Zhipeng Chen, Xiaobo Qin, Youbin Wu +4 authors
Using Pass@k as a reward in reinforcement learning with verifiable rewards improves exploration and reveals that exploration and exploitation can mutually enhance each other.
Haoyuan Wu, Haoxing Chen, Xiaodong Chen +10 authors
Grove MoE, a novel architecture with heterogeneous experts of varying sizes, improves computational efficiency and performance in large language models by dynamically activating parameters based on input complexity.
Jiejun Tan, Zhicheng Dou, Yan Yu +4 authors
HierSearch, a hierarchical agentic deep search framework using hierarchical RL, improves performance in multi-source retrieval tasks by coordinating local and Web search agents and refining knowledge.
Hasan Abed Al Kader Hammoud, Kumail Alhamoud, Abed Hammoud +3 authors
A curriculum learning strategy using Group Relative Policy Optimization (GRPO) enhances the reasoning abilities of large language models by progressively tightening token budgets, improving accuracy and token efficiency.
Yuhang Liu, Zeyu Liu, Shuanghe Zhu +10 authors
Adaptive Exploration Policy Optimization (AEPO) enhances semantic alignment in Multimodal Large Language Models (MLLMs) for GUI interaction, improving performance on benchmarks by up to 9.0% compared to RLVR.
Shuyuan Tu, Yueming Pan, Yinming Huang +6 authors
StableAvatar, an end-to-end video diffusion transformer, synthesizes infinite-length high-quality audio-driven avatar videos with natural synchronization and identity consistency using a Time-step-aware Audio Adapter and Audio Native Guidance Mechanism.
Junyan Ye, Dongzhi Jiang, Zihao Wang +9 authors
Echo-4o-Image, a synthetic dataset generated by GPT-4o, enhances image generation models by addressing rare scenarios and providing clean supervision, leading to improved performance and transferability.
Haowen Wang, Yun Yue, Zhiling Ye +12 authors
GRAO, a unified framework combining SFT and RL, enhances language model alignment through multi-sample generation, Group Direct Alignment Loss, and reference-aware parameter updates, demonstrating superior performance across human alignment tasks.
Wenhao Zeng, Yaoning Wang, Chao Hu +4 authors
ASAP, a novel coarse-to-fine framework, compresses Chain-of-Thought in code reasoning by preserving core structure and essential steps, reducing costs and improving efficiency.
Zhanghao Hu, Qinglin Zhu, Siya Qi +3 authors
Large Language Models (LLMs) have shown improved generation performance through retrieval-augmented generation (RAG) following the retriever-reader paradigm, which supplements model inputs with externally retrieved knowledge. However, prior work often evaluates RAG holistically, assessing the retriever and reader jointly, making it difficult to isolate the true contribution of retrieval, particularly given the prompt sensitivity of LLMs used as readers. We introduce Spectrum Projection Score (SPS), a lightweight, supervision-free metric that allows the reader to gauge the semantic alignment of a retrieved summary with its hidden representation by comparing the area formed by generated tokens from the summary, and the principal directions of subspace in the reader and to measure the relevance. Building on SPS we present xCompress, an inference time controller framework that dynamically samples, ranks, and compresses retrieval summary candidates. Extensive experiments on five QA benchmarks with four open source LLMs show that SPS not only enhances performance across a range of tasks but also provides a principled perspective on the interaction between retrieval and generation.
Wonjun Kang, Byeongkeun Ahn, Minjae Lee +4 authors
UNCAGE, a training-free method using contrastive attention guidance, enhances compositional fidelity in text-to-image generation by prioritizing the unmasking of object-representing tokens.
Yong Du, Yuchen Yan, Fei Tang +5 authors
GUI-RC and GUI-RCPO enhance GUI grounding accuracy by leveraging spatial consistency and reinforcement learning without additional training data.
Junjie Ye, Changhao Jiang, Zhengyin Du +7 authors
An automated pipeline for constructing training environments and a verifiable reward mechanism enhance large language models' tool-use performance without compromising general capabilities.
Vaishnavi Shrivastava, Ahmed Awadallah, Vidhisha Balachandran +3 authors
GFPO reduces length inflation in large language models by sampling larger groups and filtering responses based on length and token efficiency, maintaining accuracy and improving computational efficiency.
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