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
Shufan Jiang, Chios Chen, Zhiyang Chen
Large language models struggle with autonomous bug discovery in complex runtime environments, as demonstrated by a new game development benchmark that reveals limited effectiveness of current approaches despite sophisticated multi-agent systems and interactive agents.
Tanmay Gupta, Piper Wolters, Zixian Ma +13 authors
Open-source web agents leveraging diverse mixed datasets achieve state-of-the-art performance on browser-based tasks while operating without access to HTML or accessibility tree information.
Qisheng Su, Shiting Huang, Zhen Fang +3 authors
Researchers introduce PTE (Prefill Token Equivalents), a hardware-aware metric for measuring efficiency in Tool-Integrated Reasoning scenarios, which better correlates with actual inference latency than traditional token counts by accounting for KV-Cache inefficiencies and long tool responses.
Shilin Yan, Jintao Tong, Hongwei Xue +6 authors
Agents with meta-cognitive deficits struggle with tool usage decisions, leading to inefficiencies; a new framework called HDPO addresses this through decoupled optimization channels for accuracy and efficiency.
Shenzhi Yang, Guangcheng Zhu, Bowen Song +7 authors
Reinforcement Learning with Verifiable Rewards faces challenges with noisy labels, but a proposed method called Online Label Refinement addresses this by progressively correcting labels based on policy improvement and consistency checks, demonstrating improved robustness across various mathematical reasoning benchmarks.
Difan Jiao, Qianfeng Wen, Blair Yang +2 authors
ThinkTwice is a two-phase framework that jointly optimizes large language models for reasoning and self-refinement using Group Relative Policy Optimization, demonstrating improved performance on mathematical reasoning benchmarks.
Yujian Liu, Jiabao Ji, Li An +3 authors
Research demonstrates that skill utilization in LLM-based agents degrades significantly under realistic conditions where skills must be retrieved and refined rather than handcrafted, though targeted refinement strategies can partially restore performance.
Jianhui Liu, Haoze Sun, Wenbo Li +11 authors
OpenSpatial presents an open-source data engine for spatial reasoning tasks using 3D bounding boxes, creating a large-scale dataset and achieving state-of-the-art performance in spatial perception benchmarks.
Shuai Liu, Shulin Tian, Kairui Hu +6 authors
FileGram is a framework for personalized AI agents that uses file-system behavioral traces to enhance memory systems and agent personalization, featuring a data engine, diagnostic benchmark, and memory architecture built from atomic actions and content changes.
Nicolas Boizard, Théo Deschamps-Berger, Hippolyte Gisserot-Boukhlef +2 authors
Adapting causal generative language models into bidirectional encoders through systematic ablation and novel merging strategies achieves superior performance across multiple modalities.
InSpatio Team, Donghui Shen, Guofeng Zhang +20 authors
INSPATIO-WORLD presents a real-time framework for generating high-fidelity dynamic scenes from single videos using spatiotemporal autoregressive architecture and joint distribution matching distillation.
Ziqi Jin, Lei Wang, Ziwei Luo +1 authors
MARS is a fine-tuning method that enables autoregressive language models to predict multiple tokens per forward pass without architectural changes, maintaining accuracy while improving throughput and supporting dynamic speed adjustment.
Yuqi Zhu, Jintian Zhang, Zhenjie Wan +7 authors
LightThinker and LightThinker++ enable efficient large language model reasoning through dynamic compression and adaptive memory management, significantly reducing computational overhead while maintaining performance in complex tasks.
Haonian Ji, Kaiwen Xiong, Siwei Han +9 authors
ClawArena evaluates AI agents' ability to maintain accurate beliefs in dynamic, multi-source information environments through diverse professional scenarios and evaluation methods.
Qianshan Wei, Yishan Yang, Siyi Wang +12 authors
A new benchmark evaluates multimodal agentic capabilities by verifying tool usage and process efficiency rather than just final answers, revealing significant challenges in real-world multimodal problem-solving.
Yuxuan Zhang, EunJeong Hwang, Huaisong Zhang +8 authors
Vision-language models face challenges in video understanding due to text-based biases in benchmarks and datasets, which are addressed through VidGround, a method that uses only visually grounded questions for post-training to improve performance.
Yicheng Xiao, Wenhu Zhang, Lin Song +10 authors
A new benchmark and dataset are introduced for evaluating fine-grained spatial editing capabilities, along with a model that demonstrates superior performance on spatial manipulation tasks.
Chenxi Wang, Zhuoyun Yu, Xin Xie +8 authors
SkillX is an automated framework that creates reusable skill libraries for LLM agents through hierarchical skill design, iterative refinement, and exploratory expansion to improve generalization and efficiency across different environments.
Gabriel Sarch, Linrong Cai, Qunzhong Wang +3 authors
Vero is an open vision-language model family that achieves state-of-the-art visual reasoning performance through scaled reinforcement learning data across diverse tasks, demonstrating that broad data coverage drives strong RL scaling rather than isolated task-specific patterns.
Yitong Li, Junsong Chen, Shuchen Xue +8 authors
A novel two-stage reinforcement learning framework called Sol-RL integrates FP4 quantization with diffusion model alignment to accelerate training while maintaining high-fidelity performance.
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