arXiv：cs.AI· Andrew Zhang, Chengzhan Li·· 13 小时前
Agent Step Value：审计黑盒智能体轨迹中的评估通道敏感性
Agent Step Value: Auditing Evaluation-Channel Sensitivity in Black-Box Agent Traces
AI 导读
研究提出 Agent Step Value（ASV），以通道索引的前后目标边际增益识别状态与通道的交互作用。在冻结的 PubMed 问答转换上，48.3% 的跨通道增益符号不一致，而同通道重试仅 0.3%。
正文
Abstract:Reusing evaluator-derived step rewards across evaluation channels raises a measurement question: does the assigned direction remain stable? We define Agent Step Value (ASV) as a channel-indexed before/after target-margin gain and identify the state-by-channel interaction from a matched 2x2 design. Across frozen PubMed question-answering transitions, cross-channel gain signs disagree on 48.3% of comparisons, compared with 0.3% for same-channel retries. Transition interactions are nearly balanced in sign (median +0.014; 10% trimmed mean -0.033), but asymmetric negative magnitudes produce a positive direct mean and a negative generated-view mean under the empirical target mixture. Equal-target weighting preserves a negative mean interaction but not this strict mean reversal. Matched retrieval designs localize the negative component to the generated-view coordinate and retain its direction across three sign checks that vary readout and evaluator stack. In a matched 512-token source-contract follow-up, source-only generation restores the positive mean direction on artifact-bearing retrievals and removes parser-detected support claims from artifact-free before-state views. These results support an explanation in which generated support inflates the pre-retrieval baseline. The resulting audit can localize channel sensitivity and test an intervention before step rewards are reused.
| Comments: | Accepted at the REALM Workshop @ EMNLP 2026 (2nd Workshop for Research on Agent Language Models). 16 pages, 5 figures. The two authors contributed equally |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.04419 [cs.AI] |
| (or arXiv:2607.04419v5 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2607.04419 arXiv-issued DOI via DataCite |
Submission history
From: Chengzhan Li [view email]
[v1]
Sun, 5 Jul 2026 17:27:56 UTC (740 KB)
[v2]
Tue, 7 Jul 2026 17:26:47 UTC (50 KB)
[v3]
Mon, 13 Jul 2026 16:54:28 UTC (135 KB)
[v4]
Fri, 17 Jul 2026 17:39:00 UTC (152 KB)
[v5]
Fri, 9 Oct 2026 12:43:19 UTC (155 KB)
来源：arXiv：cs.AI · arxiv.org