Runtime policy verifiers are a common safety mechanism for AI agents: they check each action against a policy and block anything disallowed. But according to a new arXiv paper, the decisions these verifiers make are typically used only for enforcement and then thrown away. The authors ask whether those same decisions could be repurposed as a learning signal.
The paper, titled "Policy-Verified Self-Training: Learning Agent Compliance from Verifiable Policy Rewards," proposes that verifier decisions might become "verifiable policy rewards" for training. Instead of treating enforcement and learning as separate processes, the idea is to let agents improve their compliance by learning from the very checks that keep them safe.
Because the abstract is brief, the full method and experimental results are not yet clear from the announcement. But the core question is notable: it reframes safety verifiers not just as guardrails, but as a potential source of training data for more policy-compliant behavior.