Meituan Merchant Business Diagnosis via Policy-Guided Dual-Process User Simulation

The buried angle here is who benefits from cheaper counterfactual simulation: Meituan, not merchants. A system that lets a platform test pricing and promotion strategies on simulated merchant cohorts before rollout concentrates experimental power further on the platform side, with merchants having no visibility into how their behavioral profiles are being modeled.

The closest recent coverage is QuantCode-Bench, which benchmarks LLMs on generating executable trading strategies, raising a parallel question about who controls the feedback loop when AI handles strategy evaluation at scale. Both papers sit in the same emerging category: AI systems that replace costly real-world trials with learned simulations, shifting decision authority toward whoever owns the simulation. The CoopEval paper from the same date is also worth noting, since it finds LLM agents systematically defect in social dilemmas, which is a reasonable prior for how a platform-side simulation optimizing merchant policy might behave toward merchant interests when objectives diverge.

Watch whether Meituan publishes external validation showing PGHS predictions match live experiment outcomes within a stated error margin. Without that, the system's value as a merchant diagnostic tool rather than a platform optimization tool remains unverified.