4 papers
HINTBench: Horizon-agent Intrinsic Non-attack Trajectory Benchmark
Jiacheng Wang, Jinchang Hou, Fabian Wang +3
Existing agent-safety evaluation has focused mainly on externally induced risks. Yet agents may still enter unsafe trajectories under benign conditions. We study this complementary…
Token-Level Policy Optimization: Linking Group-Level Rewards to Token-Level Aggregation via Sequence-Level Likelihood
Xingyu Lin, Yilin Wen, Du Su +5
Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs), particularly in their mathemat ical reasoning performanc…
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…
Token-Level Policy Optimization: Linking Group-Level Rewards to Token-Level Aggregation via Markov Likelihood
Xingyu Lin, Yilin Wen, En Wang +4
Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs), particularly by boosting their mathematical performance.…