---
id: 20260927-T0-03
title: "独立训练的RL策略能否互相理解？可审计性研究揭示规则重叠与行为一致性"
title_en: "Independently Trained RL Policies Can Be Audited via Rule Overlap and Behavioural Agreement"
url: https://ai.daily.yangsir.net/daily/20260927-T0-03
issue_date: 2026-09-27
publish_date: 2026-09-26T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.28581
---

# 独立训练的RL策略能否互相理解？可审计性研究揭示规则重叠与行为一致性

强化学习策略通常以不透明的神经网络检查点形式分发，训练日志只记录运行过程，却不解释策略学到了什么。该研究探讨独立训练的RL策略能否被表示和比较，提出从规则重叠、行为一致性和组合性三个维度评估可审计性。核心发现是：可审计性并非单一属性，不同策略在规则层面可能存在重叠，行为上也能达成部分一致，但组合后的表现需要单独验证。这意味着未来RL系统部署时，开发者可以通过行为对比来判断黑盒策略是否遵循预期规则，而非仅依赖训练日志。

## English Version

**Independently Trained RL Policies Can Be Audited via Rule Overlap and Behavioural Agreement**

Reinforcement learning (RL) policies are often distributed as opaque neural checkpoints, with training logs showing a run occurred but not what the policy learned. This paper studies whether independently trained policies can be represented and compared, proposing that auditability is not one property but spans rule overlap, behavioural agreement, and composition. The key finding: different policies may share overlapping rules and exhibit partial behavioural agreement, but composed behaviour requires separate verification. For developers, this suggests black-box RL policies can be audited through behavioural comparison against expected rules rather than relying solely on training logs.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.28581)

**详情页**：https://ai.daily.yangsir.net/daily/20260927-T0-03

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*