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
Jie Wu, Haoling Li, Xin Zhang +7 authors
Code LLMs trained on fully synthetic data using a feature-based synthesis pipeline achieve superior performance on competitive programming benchmarks while reducing dependence on real-world coding datasets.
Jiajie Zhang, Xin Lv, Ling Feng +2 authors
A citation-aware reward framework and policy optimization method improve deep search agents' reasoning comprehensiveness and factual accuracy while reducing shortcut exploitation and hallucinations.
Siyuan Hu, Kevin Qinghong Lin, Mike Zheng Shou
Building intelligent agents capable of dexterous manipulation is essential for achieving human-like automation in both robotics and digital environments. However, existing GUI agents rely on discrete click predictions (x,y), which prohibits free-form, closed-loop trajectories (e.g. dragging a progress bar) that require continuous, on-the-fly perception and adjustment. In this work, we develop ShowUI-π, the first flow-based generative model as GUI dexterous hand, featuring the following designs: (i) Unified Discrete-Continuous Actions, integrating discrete clicks and continuous drags within a shared model, enabling flexible adaptation across diverse interaction modes; (ii) Flow-based Action Generation for drag modeling, which predicts incremental cursor adjustments from continuous visual observations via a lightweight action expert, ensuring smooth and stable trajectories; (iii) Drag Training data and Benchmark, where we manually collect and synthesize 20K drag trajectories across five domains (e.g. PowerPoint, Adobe Premiere Pro), and introduce ScreenDrag, a benchmark with comprehensive online and offline evaluation protocols for assessing GUI agents' drag capabilities. Our experiments show that proprietary GUI agents still struggle on ScreenDrag (e.g. Operator scores 13.27, and the best Gemini-2.5-CUA reaches 22.18). In contrast, ShowUI-π achieves 26.98 with only 450M parameters, underscoring both the difficulty of the task and the effectiveness of our approach. We hope this work advances GUI agents toward human-like dexterous control in digital world. The code is available at https://github.com/showlab/showui-pi.
Shuo Zhang, Chaofa Yuan, Ryan Guo +11 authors
EvoFSM is a structured self-evolving framework for LLM agents that uses finite state machines to improve adaptability while maintaining control through constrained optimization and memory mechanisms.
Lijun Liu, Linwei Chen, Zhishou Zhang +7 authors
SkinFlow introduces a novel framework for dermatological vision-language modeling that improves diagnostic accuracy through optimized visual information transmission efficiency rather than parameter scaling alone.
Xiaoshuai Song, Haofei Chang, Guanting Dong +3 authors
EnvScaler automates the creation of tool-interaction environments through programmatic synthesis, enhancing LLM performance in complex multi-turn, multi-tool tasks via supervised fine-tuning and reinforcement learning.
Xinyu Zhu, Yuzhu Cai, Zexi Liu +11 authors
ML-Master 2.0 enables long-term autonomous machine learning engineering through hierarchical cognitive caching that manages extended context and learns from execution traces.
Hengyu Shen, Tiancheng Gu, Bin Qin +10 authors
A large-scale Chinese image-text dataset called DanQing is introduced to advance vision-language pretraining, demonstrating superior performance in various downstream tasks through continual pretraining of the SigLIP2 model.
Christopher Clark, Jieyu Zhang, Zixian Ma +18 authors
Molmo2 is a new open-source video-language model family that achieves state-of-the-art performance through novel datasets and training methods, particularly excelling in video grounding tasks without relying on proprietary models.
Weili Nie, Julius Berner, Nanye Ma +3 authors
Transition Matching Distillation enables efficient video generation by distilling diffusion models into few-step predictors using conditional flows and semantic representation decomposition.
Yang-Che Sun, Cheng Sun, Chin-Yang Lin +4 authors
3AM enhances video object segmentation by integrating 3D-aware features from MUSt3R into SAM2, achieving improved viewpoint consistency with only RGB input at inference.
Delong Chen, Tejaswi Kasarla, Yejin Bang +6 authors
Action100M is a large-scale video action dataset constructed from internet instructional videos using automated pipelines with V-JEPA embeddings and GPT-based reasoning for structured annotations.
Seongyun Lee, Yongrae Jo, Minju Seo +2 authors
NoisyBench benchmark reveals significant performance degradation in state-of-the-art models when exposed to noisy contextual information, with agentic workflows amplifying errors and attention mechanisms disproportionately focusing on distractor tokens.
Siqi Kou, Jiachun Jin, Zetong Zhou +8 authors
Text-to-image diffusion models enhanced with language model reasoning capabilities achieve improved factual consistency and semantic alignment through a think-then-generate paradigm with dual-gradient reinforcement optimization.
Chengzhuo Tong, Mingkun Chang, Shenglong Zhang +12 authors
Chain-of-Frame reasoning is integrated into text-to-image generation through progressive visual refinement with explicit intermediate steps, achieving superior performance on benchmark datasets.
Tal Reiss, Daniel Winter, Matan Cohen +4 authors
Alterbute presents a diffusion-based approach for editing object intrinsic attributes while preserving identity and context through relaxed training objectives and visual named entities for scalable supervision.
Lang Feng, Fuchao Yang, Feng Chen +5 authors
AgentOCR reduces token consumption in agentic systems by representing interaction history as visual tokens and employing visual caching and self-compression techniques.
Jingsheng Zheng, Jintian Zhang, Yujie Luo +5 authors
Autonomous machine learning agents overcome execution bottlenecks by predicting outcomes before physical execution, achieving faster convergence and improved performance through a predict-then-verify approach.
Xingjun Ma, Yixu Wang, Hengyuan Xu +18 authors
Frontier language and vision models show varied safety performance across different evaluation criteria, highlighting the need for comprehensive, standardized safety assessments.
Yutao Mou, Zhangchi Xue, Lijun Li +4 authors
A guardrail model and reasoning framework are developed to detect and prevent unsafe tool invocations in LLM agents, improving both safety and task performance under adversarial conditions.
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