Learning from interaction experience is essential for enabling large language models to continually improve during deployment. Context distillation offers a promising path: privileged information extracted from experience is provided to a teacher, whose behavior is then distilled into a student model that has no access...
Zhe-Zheng Hao, Yong-Kang Yang, Xia-Ling-Zhang-Yu-Xin Xia et al.· 0 citations
DANGERLENS is introduced, a probe that maps hidden states into vocabulary space and uses blinded semantic judgments to distinguish harmful goals and operational means from hazard recognition and safety control and introduces a direct audit of intermediate representations.
Tian-Zhuo Yang, Shao-Heng Yan, Lin Lu et al.· 0 citations
This work introduces AgentBound, the first four-way counterfactual generation-and-evaluation framework for tool-using agent safety, and shows that effective agentic alignment requires action decisions to track permission-relevant execution evidence, rather than refusal strength alone.