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
Jincheng Ren, Siwei Wu, Yizhi Li +8 authors
TACO is a self-evolving compression framework that automatically discovers and refines compression rules from interaction trajectories to improve long-horizon agent performance while reducing token overhead.
Salman Rahman, Jingyan Shen, Anna Mordvina +3 authors
Research reveals that model generalization in reasoning tasks under weak supervision depends on reward saturation dynamics and reasoning faithfulness, with supervised fine-tuning on explicit traces being crucial for successful adaptation.
Jiaqing Liang, Jinyi Han, Weijia Li +15 authors
GenericAgent is a self-evolving large language model agent system that maximizes context information density through hierarchical memory, reusable SOPs, and efficient compression to overcome long-horizon limitations.
Muzhi Zhu, Shunyao Jiang, Huanyi Zheng +9 authors
Generative spatial intelligence benchmark evaluates and enhances 3D spatial constraint manipulation in image generation through real-world and synthetic datasets.
Ziao Zhang, Kou Shi, Shiting Huang +13 authors
SkillFlow presents a benchmark for evaluating autonomous agents' ability to discover, repair, and maintain skills over time through a structured lifelong learning protocol.
Yi Zhong, Buqiang Xu, Yijun Wang +4 authors
Chat2Workflow presents a benchmark and agentic framework for automating executable visual workflow generation from natural language, revealing significant challenges in achieving industrial-grade automation despite advances in language models.
Jiaxi Bi, Tongxu Luo, Wenyu Du +2 authors
STOP is a systematic path pruning method for large reasoning models that improves efficiency and accuracy through learnable token-level pruning across different compute budgets.
Xiyang Wu, Zongxia Li, Guangyao Shi +5 authors
A co-evolution framework enables large language models to discover, retain, and reuse structured skills across episodes in long-horizon interactive environments through a learnable skill bank and skill pipeline.
Chaitanya Dwivedi, Binxuan Huang, Himanshu Gupta +3 authors
Expert upcycling expands Mixture-of-Experts capacity during continued pre-training by duplicating experts and extending routers while maintaining fixed inference cost, achieving better training efficiency and model quality.
Xinping Lei, Xinyu Che, Junqi Xiong +16 authors
WebCompass evaluates web development capabilities through diverse input modalities and task types, using automated evaluation methods that simulate real-world coding workflows.
Constantinos Karouzos, Xingwei Tan, Nikolaos Aletras
Output diversity collapse in post-trained language models is primarily driven by training data composition rather than generation format, with different post-training methods affecting diversity differently across tasks.
Junghwan Lim, Wai Ting Cheung, Minsu Ha +25 authors
Motif-Video 2B achieves high text-to-video generation quality using a specialized architecture with shared cross-attention and three-part backbone, along with efficient training methods, while requiring significantly fewer parameters and training data than larger models.
Qijun Han, Haoqin Tu, Zijun Wang +11 authors
VLAA-GUI is a modular GUI agent framework that addresses early stopping and repetitive loop issues through integrated components for verification, loop breaking, and search capabilities.
Yen-Siang Wu, Rundong Luo, Jingsen Zhu +6 authors
Video speed manipulation and perception models are developed through self-supervised temporal reasoning, enabling speed detection, slow-motion video generation, and temporal super-resolution from in-the-wild sources.
Joachim Baumann, Vishakh Padmakumar, Xiang Li +3 authors
SWE-chat presents a large-scale dataset of real coding agent interactions that reveals significant inefficiencies and challenges in current AI-assisted development practices.
Gene Chou, Charles Herrmann, Kyle Genova +6 authors
CityRAG generates long-term, physically grounded video sequences that maintain environmental consistency and support complex navigation through real-world geography using geo-registered data as context.
Yixuan Tang, Yi Yang
LoRA adapter merging performance can be improved by separately calibrating the output-side matrix B to reduce interference from shared directions while preserving task-specific information.
Wentao Shi, Yu Wang, Yuyang Zhao +8 authors
Agent-as-a-Judge benchmark evaluates automated verification capabilities across multiple domains with comprehensive task assessment.
Shaden Alshammari, Kevin Wen, Abrar Zainal +5 authors
MathNet is a large-scale, multilingual, multimodal dataset of Olympiad-level math problems designed for evaluating mathematical reasoning and retrieval in generative models and embedding-based systems.
Alexey Khoroshilov, Alexey Chernysh, Orkhan Ekhtibarov +2 authors
QuantCode-Bench evaluates large language models on generating executable trading strategies by testing their ability to translate natural language descriptions into functional code that operates correctly on historical financial data.
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