🛰️ Daily AI Frontier
‹ back to 2026-08-21

PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents

Research LLM Agents

Ranking

Overall 86
Content 95
Popularity 65

Observed public metrics from 1 member.

Merged summary

TL;DR - PolicyGuide compiles organizational policies into workflow graphs and proactively guides customer-service LLM agents through compliant multi-step procedures. It improves both action-level safety and procedural compliance across several service domains and model families.

  • At user-turn boundaries, a verifier uses persisted workflow state to reconcile open requests and recommend step-specific remediation.
  • On τ²-bench airline, retail, and telecom tasks, PolicyGuide increased mean Pass⁴ from 0.42 to 0.62 with GPT-5.4.
  • Telecom saw the largest improvement, from 0.19 to 0.61, reflecting its highly structured workflows.
  • The workflows transferred to Claude Sonnet 4.6 and Gemini 2.5 Pro and achieved the lowest observed adversarial attack-success rate.

Sources (1)

PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents

arXiv cs.AI Seongjae Kang, Taehyung Yu, Sung Ju Hwang 2026-08-20 arXiv:2608.19861
Public signals Hugging Face upvotes 9
Providers: Hugging Face · Upvotes 9 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-19 14:25:39.917379 UTC

TL;DR - PolicyGuide compiles organizational policies into workflow graphs and proactively guides customer-service LLM agents through compliant multi-step procedures. It improves both action-level safety and procedural compliance across several service domains and model families.

  • At user-turn boundaries, a verifier uses persisted workflow state to reconcile open requests and recommend step-specific remediation.
  • On τ²-bench airline, retail, and telecom tasks, PolicyGuide increased mean Pass⁴ from 0.42 to 0.62 with GPT-5.4.
  • Telecom saw the largest improvement, from 0.19 to 0.61, reflecting its highly structured workflows.
  • The workflows transferred to Claude Sonnet 4.6 and Gemini 2.5 Pro and achieved the lowest observed adversarial attack-success rate.
item →