这篇论文教你怎么用 ASV 框架来拆解 AI agent 每一步的决策价值，在开放问答任务上实测发现短 rationale 反而有害，做 agent 评估的必看。
论文提出 Agent Step Value (ASV) 框架，通过重放黑盒轨迹并用无状态 LLM 评估器打分，量化每一步的价值。在 100 道开放问答任务（使用 PubMed 实时检索和 DeepSeek 对数概率评分）上，ASV 评估了 1,100 次转移，结果显示熵移动为 0.000，平均贝叶斯惊讶为 2.693，表明信念近乎 one-hot 转折。在 128 个 token 的 rationale 条件下，平均黄金边际增益为 -2.335（95% CI [-3.395, -1.272]）；而直接用单 token 评分同一轨迹得到 +4.033。论文通过 100 次转移的组件审计追踪到反转源于短 rationale 对完整状态的覆盖。
Agent Step Value: Probing the Observer Effect in Black-Box Traces
Final-answer scores hide which agent transitions helped or harmed a trace. We introduce Agent Step Value (ASV), a replay framework that scores before/after states with a stateless LLM evaluator over a fixed candidate set. ASV reports entropy movement and Bayesian surprise measure belief movement, while offline gold-margin gain measures movement toward a reviewed target. It also quantifies evaluator-channel sensitivity by replaying the same frozen transitions under changed projection, rationale, prompt, or scoring rules. In a 100-question open-QA study with live PubMed retrieval and DeepSeek log-probability scoring, ASV evaluates 1,100 transitions. Entropy movement is 0.000 while mean Bayesian surprise is 2.693, exposing near-one-hot belief pivots. Under a 128-token rationale-conditioned protocol, mean gold-margin gain is -2.335 (95\% CI [-3.395, -1.272]); direct one-token scoring on the same traces gives +4.033. A 100-transition component audit traces the reversal to short generated rationales over full states. ASV turns prompt sensitivity into a measured channel effect and localizes the largest rationale-conditioned losses to extraction and audit.