arXiv · Computation and Language· Jiayi Li, Ruizhe Li·· 6 小时前AI 评分39
When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs
When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs
AI 导读
summary_zh: SAKIKO 审计框架通过定向误差发现、路由器条件干预、目的地验证和统计许可冻结，形式化表示修复。在 When2Call 和 MetaTool 上，七个 LLM 中五个模型获得方向特异性净增益；59 个预算匹配随机方向无一达到校准目标增益。
来源：arXiv · Computation and Language · arxiv.org