arXiv cs.AI· Houcheng Jiang, Boxuan Zhang, Qiyong Zhong, Junfeng Fang, Xiang Wang, Xiangnan He·· 14 小时前AI 评分34
RePolicy：让 AI 智能体安全防护学会选择并应用安全策略
RePolicy: Learning to Select and Apply Safety Policies for Agent Safeguards
AI 导读
RePolicy 是一种联合学习策略选择与安全评估的智能体安全防护方案，可根据智能体轨迹从候选策略中选出最匹配的一项并判断行为是否安全。研究团队构建了含超 2 万条智能体轨迹的 PolicyTraj-20K 数据集，采用监督示范加强化学习训练，并通过随机化策略顺序、标识符和干扰项减少捷径学习。
正文 · 原文
Abstract:Agent safeguards must determine whether an agent's behavior violates the safety rules of the current task, environment, and user. Because these rules vary across contexts, a safeguard needs to select the relevant policy before judging the behavior. Existing policy-aware safeguards can apply a supplied policy, but generally assume that the correct policy is already known. They therefore do not learn how to select a policy from a library, even though an incorrect choice can lead to an incorrect safety judgment. We introduce RePolicy, a safeguard that jointly learns policy selection and safety assessment. Given an agent trajectory and a set of candidate policies, RePolicy selects the policy that best matches the situation, applies its requirements, and predicts whether the behavior is safe. We construct PolicyTraj-20K, a dataset of more than 20,000 agent trajectories annotated with applicable policies, safety labels, and explanations. We train RePolicy with supervised demonstrations followed by reinforcement learning, while randomizing policy order, identifiers, and distractors to reduce shortcut learning. Across six agent-safety benchmarks and 19 baselines, RePolicy-4B achieves leading overall safety-detection performance and improves policy-selection accuracy. These results show that safeguards can learn which safety requirements to use for assessing agent behavior.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2608.24275 [cs.AI] |
| (or arXiv:2608.24275v3 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.24275 arXiv-issued DOI via DataCite |
Submission history
From: Houcheng Jiang [view email]
[v1]
Tue, 25 Aug 2026 09:01:33 UTC (4,146 KB)
[v2]
Thu, 27 Aug 2026 06:48:15 UTC (4,145 KB)
[v3]
Thu, 29 Oct 2026 10:16:40 UTC (2,681 KB)
来源：arXiv cs.AI · arxiv.org