DigestAI news desk

Cut through the AI noise.

Research

Researchers propose Policy-as-Skill for governed LLM decision-making

Researchers introduced Policy-as-Skill (PaS), a modular runtime framework designed to integrate governance, evidence validation, and auditability into large language model (LLM) decision-making. The system packages policy functions as executable, versioned capabilities to handle tasks like compliance checks, risk assessments, and operational decisions.

1 source primary source

Key points

  • Policy-as-Skill (PaS) adds governance, evidence checks, and audit trails to LLM decision-making via modular, versioned policy modules
  • Tested on 600 tasks with **Gemma4**, achieving **53.8% exact accuracy** and **0.346 macro-F1** without deterministic control
  • Deterministic control boosts accuracy to **61.2%** but works best for specific tasks, not universally

Evaluated on 600 development tasks using a fixed Gemma4 backend, PaS+Audit achieved 53.8% exact accuracy and macro-F1 of 0.346. With deterministic control, accuracy rose to 61.2%, though performance varied by task. The framework outperformed traditional LLM+RAG approaches on governance and review metrics, including citation precision of 1.000 and policy-reference recall of 0.984. The authors emphasize its potential for selective rule-based intervention rather than universal deterministic control.

Read the original at arXiv cs.AI · by Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya primary sourceOpen source ↗
Topics · follow one to build your own front page

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.

Comments

via GitHub Discussions

More in Research

All →

Related stories