U.S. FEDResearch
Code-as-Auditor paper maps regulations into executable LLM compliance checklists
Researchers led by Jisoo Kim posted Code-as-Auditor to arXiv on September 16, 2026, proposing an LLM framework that turns regulatory text into formalized checklists and executable decision trees for evidence-grounded compliance assessment. The CIKM 2026-accepted paper reports experiments on privacy and data-protection scenarios in which the system expands checklist items into factual and counterfactual questions and adds a self-verification loop for traceability.
Key points
- Framework translates regulations into checklists and executable decision trees expressed as interpretable code.arXiv
- Each checklist item expands into factual and counterfactual questions over case-specific evidence.arXiv
- A self-verification loop targets logical consistency of generated code and outcome traceability.arXiv
- Authors tested the approach on privacy and data-protection compliance scenarios.arXiv
- Paper was accepted at ACM CIKM 2026 in Rome (7–November 11, 2026).arXiv