As LLM agents advance rapidly in autonomy, the risk of hidden objectives that bypass safety constraints becomes increasingly plausible. Chain-of-thought (CoT…
机构:清华
来源:arXiv 2611.07586 | AI4Papers 论文推荐平台