04:00
2026-07-07
arxiv.org
artificial-intelligence
When Aggregate Alignment Misleads: Auditing Policy Repair Without Per-State Expert Actions
A new study on agentic AI policy repair in a hotel-pricing simulator shows that LLM-based editors can nearly match benchmark performance using only region-level diagnostic feedback, achieving RevPAR 1β¦