5 papers
Not All Tokens Learn Alike: Attention Entropy Reveals Heterogeneous Signals in RL Reasoning
Gengyang Li, Zheng-Fan Wu, Siqi Bao +1
Reinforcement-learning-based post-training has become a key approach for improving the reasoning ability of large language models, but its token-level learning signals remain poorl…
Securing the Floor and Raising the Ceiling: A Merging-based Paradigm for Multi-modal Search Agents
Zhixiang Wang, Jingxuan Xu, Dajun Chen +3
Recent advances in Vision-Language Models (VLMs) have motivated the development of multi-modal search agents that can actively invoke external search tools and integrate retrieved…
Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis
Wang Cai, Yilin Wen, Jinchang Hou +5
Safety alignment in Large Language Models (LLMs) inherently presents a multi-objective optimization conflict, often accompanied by an unintended degradation of general capabilities…
SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation
Gengyang Li, Wang Cai, Yifeng Gao +1
Chain-of-Thought (CoT) prompting improves reasoning but often produces long and redundant traces that substantially increase inference cost. We present SyncThink, a training-free a…
ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy
Gengyang Li, Yifeng Gao, Yuming Li +1
While Chain-of-Thought (CoT) prompting improves reasoning in large language models (LLMs), the excessive length of reasoning tokens increases latency and KV cache memory usage, and…