1
Counterfactual Trace Auditing of LLM Agent Skills
反事实追踪审计方法，可揭示LLM智能体技能的可解释性与可靠性缺陷，提升AI安全。
arXiv:2605.11946v2 Announce Type: replace Abstract: Large Language Model agents are increasingly augmented with agent skills. Current evaluation metho…