PolicyGuard: A Dialogue-Grounded Sub-Agent Verifier for Policy Adherence in LLM Agents

Official implementation of PolicyGuard, a dialogue-grounded sub-agent verifier that enforces policy adherence in LLM agents by sharing the full conversation context and providing reasoning-based corrective feedback — rather than isolated argument-level blocking.

Evaluated on TAU-bench, PolicyGuard achieves up to +12 pp task success improvement across multiple LLM backbones while cutting false-positive blocks by roughly half compared to argument-level guards.

🔜 Coming Soon

Code will be released soon.

📝 Citation

@article{kang2026policyguard,
  title={PolicyGuard: A Dialogue-Grounded Sub-Agent Verifier for Policy Adherence in LLM Agents},
  author={Kang, Seongjae and Yu, Taehyung and Hwang, Sung Ju},
  journal={arXiv preprint arXiv:2606.29225},
  year={2026}
}