mlsanityData to 7 Oct 2026

Agent Step Value: Auditing Evaluator-Channel Reversals in Black-Box Agent Traces

arXiv 2607.04419 ↗First posted 5 Jul 2026cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

Pooling, substituting, or reusing evaluator-derived step rewards assumes that their direction survives a change of evaluation channel. The same frozen transition can violate that assumption. Process rewards vary agent states, while evaluator audits vary scoring configurations; neither first difference isolates their interaction. We define Agent Step Value (ASV) as a channel-indexed target-margin gain and identify the state-by-channel interaction on complete matched faces. Across frozen PubMed question-answering transitions, direct scoring yields a positive mean ASV, while the generated-view channel yields a negative mean. Two matched replay waves reproduce this reversal, and cross-channel sign disagreement exceeds same-channel retry disagreement by 48.0 percentage points. Matched retrieval faces localize the reversal to the generated-view coordinate and trace its direction across a readout-and-stack bridge. A source-only generation contract restores the positive mean direction on artifact-bearing retrievals and removes parser-detected substantive support claims from artifact-free before-state views. ASV turns channel sensitivity into an identified measurement problem that can be localized and tested by intervention before step rewards are reused.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 22 of its 41 references that are in the map. First posted on arXiv 5 Jul 2026.

Builds on

22 of 41 references.

References · 22

  1. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  2. 2023Let's Verify Step by Stepcited by 3,598
  3. 2023G-Eval: NLG Evaluation using GPT-4 with Better Human Alignmentcited by 2,205
  4. 2023Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Promptingcited by 1,081
  5. 2023Large Language Models are not Fair Evaluatorscited by 1,050
  6. 2022Solving math word problems with process- and outcome-based feedbackcited by 871
  7. 2024LLM Evaluators Recognize and Favor Their Own Generationscited by 651
  8. 2023Prometheus: Inducing Fine-grained Evaluation Capability in Language Modelscited by 558
  9. 2025The Lessons of Developing Process Reward Models in Mathematical Reasoningcited by 402
  10. 2019Measurement and Fairnesscited by 299
  11. 2025AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectoriescited by 99
  12. 2025Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learningcited by 68
10 more
  1. 2025Position: Evaluating Generative AI Systems Is a Social Science Measurement Challengecited by 61
  2. 2026AgentProcessBench: Diagnosing Step-Level Process Quality in Tool-Using Agentscited by 28
  3. 2024Evaluating the Consistency of LLM Evaluatorscited by 24
  4. 2025Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluationcited by 23
  5. 2025Flaw or Artifact? Rethinking Prompt Sensitivity in Evaluating LLMscited by 23
  6. 2025Scaling Evaluation-time Compute with Reasoning Models as Evaluatorscited by 13
  7. 2025ReliableEval: A Recipe for Stochastic LLM Evaluation via Method of Momentscited by 10
  8. 2026Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judgescited by 8
  9. 2026Efficient Process Reward Modeling via Contrastive Mutual Informationcited by 4
  10. 2026Beyond Math and Code: Lightweight Corpus-Grounded Process Rewards for Factual Question Answeringcited by 4

Built on by

Not cited yet.