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

2026-04-16 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Meituan has developed Policy-Guided Hybrid Simulation (PGHS), a dual-process framework designed to simulate group-level user behavior for evaluating merchant strategies without expensive online experiments. PGHS addresses two key challenges in user simulation: information incompleteness, which leads reasoning-based simulators to over-rationalize, and mechanism duality, which requires capturing both explicit preferences and implicit statistical regularities. The framework mines transferable decision policies from behavioral trajectories, using them as a shared alignment layer. This layer anchors an LLM-based reasoning branch to prevent over-rationalization and an ML-based fitting branch to absorb implicit regularities. The group-level predictions from both branches are fused for complementary correction. Deployed on Meituan with 101 merchants and over 26,000 trajectories, PGHS achieved a group simulation error of 8.80%, outperforming the best reasoning-based and fitting-based baselines by 45.8% and 40.9% respectively.

Key takeaway

For research scientists developing user behavior simulators, PGHS offers a robust framework to overcome limitations of purely reasoning-based or fitting-based approaches. You should consider integrating policy-guided dual-process simulation to achieve higher accuracy in counterfactual evaluations, especially when dealing with complex user behaviors influenced by both explicit preferences and implicit factors.

Key insights

A dual-process simulation framework improves user behavior prediction by combining LLM reasoning with ML fitting.

Principles

• Combine reasoning and fitting for robust simulation.
• Policy guidance aligns diverse simulation models.

Method

PGHS mines decision policies from trajectories, using them to align an LLM-based reasoning branch and an ML-based fitting branch, then fuses their group-level predictions.

In practice

• Use policy guidance to anchor LLM and ML models.
• Fuse diverse model outputs for error correction.

Topics

• Policy-Guided Hybrid Simulation
• User Behavior Simulation
• Merchant Business Diagnosis
• LLM-based Reasoning
• Machine Learning Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Mental-R1: Aligning LLM Reasoning for Mental Health Assessment — Aligning LLM reasoning with human cognitive processes via specialized reinforcement learning improves mental health assessment accuracy.
• BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces — BehaviorBench offers a real-world benchmark for personalized decision modeling, revealing personalization's varied impact across prediction tasks.
• GRPO-VPS: Enhancing Group Relative Policy Optimization with Verifiable Process Supervision for Effective Reasoning — Verifiable process supervision refines LLM reasoning by providing targeted, segment-wise feedback, improving GRPO's efficiency.
• Pushing Forward Pareto Frontiers of Proactive Agents with Behavioral Agentic Optimization — BAO optimizes proactive LLM agents for both task performance and user engagement using behavioral reinforcement learning.
• Multi-Agent Framework for Audit Risk Assessment with Explicit Uncertainty and Evidence Conflict Modeling — UMAR integrates multi-agent risk scores with Dempster-Shafer theory to quantify uncertainty and evidence conflict in audit risk assessment.
• Controllable and Verifiable Process Data Synthesis for Process Reward Models — Synthesized, verifiable process data with controlled errors significantly improves PRM performance in reasoning tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.