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
Yang Xiao, Yusong Sun, Haoyi Wu +17 authors
Vela converts real-world user behavior records into training tasks and trajectories, improving large language model performance on user-related and general benchmarks through supervised fine-tuning and reinforcement learning.
Xinge Peng, Zhen Fang, Ruibing Wang +8 authors
ExeWorldBench evaluates executable world modeling by assessing both structural understanding and code-based 3D construction across mechanism, compositional, and intent-driven tasks.
Xialie Zhuang, Zunhai Su, Jia Pan +3 authors
StartLux-Decision introduces efficient and adaptive decision models using single-pass prediction and confidence-guided reasoning, trained via a three-stage pipeline with RLCD² reinforcement learning, and evaluates them on Decision Index and the DecisionPivot benchmark.
Zefeng He, Yunkai Chai, Xiaoyang Chen +5 authors
LatentLoop improves looped Mixture-of-Experts by compressing the first loop visit, reallocating saved computation to larger hidden dimensions, and using separate attention parameters, yielding lower validation loss and better benchmark scores.
Xin Li, Hao Jiang, Jinghao Guo +2 authors
ReMaRk improves reinforcement learning for recommendation by scoring generated lists against all valid replies jointly, reducing training cost while matching or exceeding prior performance.
Dongxu Yang, Hongda Yan, Renxin Wang +23 authors
FDE-Gym evaluates AI agents on realistic customer delivery tasks requiring clarification of ambiguous requirements and discovery of unstated constraints, revealing a substantial performance gap compared to well-specified benchmarks.
Tianyi Wu, Chao Ge, Peng Chen +12 authors
The Last Secure Code Benchmark reveals that frontier AI coding agents generate functional but rarely secure code across diverse languages and vulnerability classes.
Qingchuan Ma, Yuexiao Ma, Xiang Meng +7 authors
Vision-language models struggle to convert 2D visual observations into precise 3D rotational control, showing large gaps versus humans in projection alignment tasks.
Jie Huang, Yanan Chen, Ruixun Liu +6 authors
EmbodiedCLBench evaluates continual learning in embodied agents by testing compositional skill reuse across basic and complex tasks, revealing stable but non-scaling experience benefits.
Fanxu Meng, Taosong Fang
DoSA accelerates sparse attention by using a certificate to conditionally score only eligible tokens, preserving exact selection while reducing computation.
Jintian Zhang, Feng He, Xinnian Liang +5 authors
Cortex consolidates context-induced improvements into mergeable low-rank or structural updates, preserving non-target capabilities across sequential edits and diverse model scales.
Jungwoo Park, Taewhoo Lee, Jaewoo Kang
UltraBERT improves bidirectional encoder performance by optimizing masking ratios, compute allocation, and distillation to achieve state-of-the-art results on both classification and retrieval benchmarks.
Yuan Feng, Qize Yang, Ruizhe Chen +9 authors
VisionWeave introduces elastic visual representation weaving to adaptively allocate visual tokens by granularity, reducing token usage by 43% while preserving nearly all model performance across diverse tasks and serving conditions.
Kaiyi Huang, Wangbo Yu, Yukun Huang +6 authors
Multics introduces hierarchical group-level communication and dynamic grouping to scale multi-player world models efficiently while reducing interference and maintaining consistency.
Hanmo Chen, Chengcheng Liu, Tianxiao Chen +7 authors
StereoBind binds visual motion to stereo sound generation via motion tracks and specialized encoders, improving spatial alignment in joint video-audio synthesis.
Yipeng Gao, Ziyao Zeng, Jinfa Huang +16 authors
3DCodeVerse is a research ecosystem that provides datasets, automated verification harnesses, and empirical guidance to enable scalable, editable, and verifiable 3D content generation via executable code.
Yilong Xu, Minrui Xu, Xinyu Zhou +3 authors
Multi-session agentic reinforcement learning is introduced to handle trajectories across multiple sessions, with Mosaic optimizing session-level contributions and allocation via hierarchical advantage estimation.
Qi Lu, Jiarun Li, Ce Zhang +5 authors
VLA-Δ improves robot control by explicitly supervising physical state transitions between current and future observations, achieving strong benchmark and real-world results without added inference cost.
Huining Yuan, Xiangmin Yi, Zelai Xu +5 authors
ActionOPD aligns smaller language-model agents with teachers via semantic action-level on-policy distillation, improving out-of-distribution tool-use performance without requiring shared tokenizers or teacher probabilities.
Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh +3 authors
Tev-Gym provides verified long-horizon training environments with dense progressive rewards, and progress-potential reward shaping enables efficient training of Tev-27B, which achieves competitive long-horizon agent 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