Medi-Sim AI simulator designs healthcare policies that eliminate up-coding and halve rejections

Current healthcare AI benchmarks fail to account for how providers strategically respond to payment rules, leading to unintended consequences like up-coding and patient selection. A new paper from researchers Zihan Wang, Xiang Xu, Hongyuan Zha, and Wenhao Li recasts hospital mechanism design as program synthesis for language models. Their system, Medi-Sim, is a multi-agent simulator that models five strategic provider channels: coding, selection, delay, effort, and triage. Using typed, inspectable rule programs scored by the simulator, an LLM-guided evolutionary code search automatically discovers optimal policies. The incentive sweep recovers classical health-economics findings: up-coding and low-complexity patient selection under profit pressure, and Goodhart-style drift where measured performance becomes anti-correlated with true outcomes.

Medi-Sim also exposes pressure migration — closing the coding channel more than doubles low-complexity selection. The LLM-guided search over the same rule-program space synthesizes an inspectable mixed-objective program that eliminates up-coding entirely, halves patient rejection rates, and retains most of the profit-oriented baseline's funds. This demonstrates that AI can design transparent, incentive-compatible healthcare mechanisms without sacrificing efficiency. The work bridges economics, multi-agent AI, and program synthesis, offering a practical tool for policymakers and hospital administrators to audit and revise payment rules before deployment.