跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2606.05263 · v2

Learning Which Evidence-Seeking Decisions Matter: Calibrated Counterfactual Replay for Language Agents

首次提交 2026-06-04 00:19
版本更新 2026-11-19 16:52
雷达观测 2026-11-20 19:58
arXiv ↗PDF ↗

摘要

Language agents answer questions by searching, reading, and checking claims. A reward on the final answer says little about which of those decisions earned it, since an agent can reach the right string after redundant searches or with half the evidence missing. We treat the question as an experiment. At a saved decision point, PCCC-RL, for prefix-conditioned counterfactual credit, restores the tool state, runs the action the agent chose alongside a feasible alternative, and lets the same frozen policy finish both branches. The difference in final score then samples what the decision was worth. Replaying everything is too expensive, so the method predicts each action's credit up front, checks a random subset, and uses the recorded selection probabilities to correct the errors of that prediction, spending the budget where a further check can still move the training signal. Alternatives are drawn from the prefix alone, which keeps the comparison a usable gradient baseline. On HotpotQA and MuSiQue, joint exact match, requiring a correct answer together with complete supporting evidence, reaches 56.71% and 36.82%, gains of 4.19 and 3.86 points over a baseline with the same replay access. The gain concentrates in complete evidence chains and accumulates over many updates. Single-update effects stay uncertain, and citation coverage and scope distinctions such as "none" versus "not all" remain open.

作者备注:31 pages, 9 figures. Accepted at the Asian Conference on Machine Learning (ACML 2026)

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-19 16:52 · 最新观测版本

Learning Which Evidence-Seeking Decisions Matter: Calibrated Counterfactual Replay for Language Agents

Language agents answer questions by searching, reading, and checking claims. A reward on the final answer says little about which of those decisions earned it, since an agent can reach the right string after redundant searches or with half the evidence missing. We treat the question as an experiment. At a saved decision point, PCCC-RL, for prefix-conditioned counterfactual credit, restores the tool state, runs the action the agent chose alongside a feasible alternative, and lets the same frozen policy finish both branches. The difference in final score then samples what the decision was worth. Replaying everything is too expensive, so the method predicts each action's credit up front, checks a random subset, and uses the recorded selection probabilities to correct the errors of that prediction, spending the budget where a further check can still move the training signal. Alternatives are drawn from the prefix alone, which keeps the comparison a usable gradient baseline. On HotpotQA and MuSiQue, joint exact match, requiring a correct answer together with complete supporting evidence, reaches 56.71% and 36.82%, gains of 4.19 and 3.86 points over a baseline with the same replay access. The gain concentrates in complete evidence chains and accumulates over many updates. Single-update effects stay uncertain, and citation coverage and scope distinctions such as "none" versus "not all" remain open.

查看此版本解读 →查看此版本原文 ↗