Safeguarding language agents requires assessing trajectories under context-dependent safety policies. Existing methods relying on prompting or supervised fine-tuning struggle to adapt to unseen trajectories and changing policy contexts. RePolicy proposes a approach that learns to invoke the correct policy from a dynamic library given an agent trajectory, producing a policy-grounded rationale and safety judgment. It uses GRPO with verifiable rewards and policy-context perturbation to improve robustness. The method is initialized using the constructed PolicyTraj-20K dataset. Evaluated on six benchmarks, RePolicy achieves strong overall safety detection and robust policy invocation across varying policy contexts. This is interesting because it directly addresses the challenge of adapting to new policies and scenarios without retraining.
Reinforcement learning strengthens policy invocation for agent safety judgments
RePolicy selects applicable rules from a changing policy library, then uses them to assess complete agent trajectories.
Top University
Houcheng Jiang · Boxuan Zhang · Qiyong Zhong · Junfeng Fang · Xiang Wang · Xiangnan He
Zhejiang University · Zhongguancun Academy · University of Science and Technology of China · National University of Singapore
Research Digest··2 min read
Thread:Agent Rule Compliance
In one line
RePolicy learns to invoke safety policies for agent safeguards via reinforcement learning, achieving strong detection performance.
Before you read this
Three ideas this paper builds onagent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.144 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.239 papers →group relative policy optimizationA reinforcement learning method that optimizes policy by computing advantages based on group-level reward statistics.154 papers →
Read it as
Jiang et al. train an agent safeguard to identify and invoke the safety policy relevant to a full execution trajectory.
The authors developed RePolicy, a safeguard that receives an agent’s execution trajectory alongside a dynamic library of safety policies.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code