What they did

The authors developed RePolicy, a safeguard that receives an agent’s execution trajectory alongside a dynamic library of safety policies. It selects the applicable policy, cites its content in a rationale, and produces a safety judgment grounded in that policy.

Training begins with supervised initialization on PolicyTraj-20K, a dataset of roughly 20,000 policy-and-trajectory examples. The model is then optimized using group relative policy optimization (GRPO), verifiable rewards, and perturbations to the policy context intended to prevent reliance on a fixed library arrangement.

Key findings

• Across six agent-safety benchmarks, RePolicy achieved strong overall safety-detection performance, according to the authors.
• The system reliably invoked applicable policies even when the surrounding policy context varied.
• Reinforcement learning with verifiable rewards supported both the final safety judgment and the intermediate task of selecting the relevant rule.
• Policy-grounded rationales linked judgments to the invoked policy rather than relying only on an implicit model assessment.

Why it matters

Agent safeguards must evaluate sequences of actions under rules that may differ by application or change over time. RePolicy treats retrieval and application of the correct policy as a learned decision problem, offering a route beyond safeguards trained only on fixed prompts or supervised examples.

Caveats

The supplied abstract does not report benchmark-level scores, baseline comparisons, computational costs, or results from real-world deployment. Robustness across six benchmarks therefore does not establish reliability under arbitrary policies, adversarially written rules, ambiguous policy conflicts, or substantially longer and more complex agent trajectories.