arXiv:2604.15190v1 [cs.AI] 16 Apr 2026
Meituan Merchant Business Diagnosis via Policy-Guided Dual-Process User Simulation Ziyang Chen
Renbing Chen
Daowei Li
LongCat Interaction Team, Meituan Inc. Beijing, China [email protected]
LongCat Interaction Team, Meituan Inc. Beijing, China [email protected]
LongCat Interaction Team, Meituan Inc. Beijing, China [email protected]
Jinzhi Liao
Jiashen Sun Ke Zeng
Xiang Zhao∗
Independent Researcher Changsha, China [email protected]
[email protected] [email protected] LongCat Interaction Team, Meituan Inc. Beijing, China
Abstract Simulating group-level user behavior enables scalable counterfactual evaluation of merchant strategies without costly online experiments. However, building a trustworthy simulator faces two structural challenges. First, information incompleteness causes reasoning-based simulators to over-rationalize when unobserved factors such as offline context and implicit habits are missing. Second, mechanism duality requires capturing both interpretable preferences and implicit statistical regularities, which no single paradigm achieves alone. We propose Policy-Guided Hybrid Simulation (PGHS), a dual-process framework that mines transferable decision policies from behavioral trajectories and uses them as a shared alignment layer. This layer anchors an LLM-based reasoning branch that prevents over-rationalization and an ML-based fitting branch that absorbs implicit regularities. Group-level predictions from both branches are fused for complementary correction. We deploy PGHS on Meituan with 101 merchants and over 26,000 trajectories. PGHS achieves a group simulation error of 8.80%, improving over the best reasoning-based and fitting-based baselines by 45.8% and 40.9% respectively.
CCS Concepts • Information systems → Information systems applications.
Keywords Merchant Business Diagnosis, User Behavior Simulation, Large Language Models ∗ Corresponding author.
This work is licensed under a Creative Commons Attribution-NonCommercialNoDerivatives 4.0 International License. SIGIR ’26, Melbourne, VIC, Australia. © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2599-9/2026/07 https://doi.org/10.1145/3805712.3808505
Independent Researcher Changsha, China [email protected]
ACM Reference Format: Ziyang Chen, Renbing Chen, Daowei Li, Jinzhi Liao, Jiashen Sun, Ke Zeng, and Xiang Zhao. 2026. Meituan Merchant Business Diagnosis via PolicyGuided Dual-Process User Simulation. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3805712.3808505
1
Introduction
Merchant business diagnosis on large-scale platforms requires answering counterfactual questions about operational strategies [6, 11, 18, 19]. For example, platform operators need to estimate whether adjusting pricing, restructuring menus, or redesigning promotional pages would improve a merchant’s conversion rate before committing to costly changes. Since each merchant serves a distinct user population with heterogeneous preferences, such diagnosis demands accurate simulation of group-level user behavior under hypothetical interventions. Online A/B testing remains the gold standard for causal evaluation [22, 23], but it suffers from critical scalability limitations. Deploying unproven strategies risks user experience degradation and financial loss [14, 20]. Modern strategies also involve high-dimensional semantic variables such as promotional copywriting and visual layouts [7, 10], creating a combinatorial space that physical experiments cannot cover [3]. These constraints motivate simulation-based counterfactual evaluation [14]. While individual choices are stochastic, the collective preferences of homogeneous user segments exhibit stable regularities, making group-level simulation a natural granularity for merchant diagnosis. Building a trustworthy group-level simulator faces two structural challenges. The first is information incompleteness. Real decisions are influenced by unobserved factors such as offline context and implicit habits, yet reasoning-based simulators rely on explicit inputs and tend to produce over-rationalized predictions [1, 21]. User simulation has evolved from rule-based and statistical approaches [2, 8] to LLM-driven agents [9, 24], but these agents often behave as hyperrational decision-makers that fail to replicate bounded rationality.
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia.
Ziyang Chen et al.
Figure 1: Overview of the PGHS framework. Phase A abstracts decision policies through macro-level persona clustering and micro-level rationale refinement. Phase B produces predictions via semantic reasoning and statistical fitting, both conditioned on the abstracted policies. Phase C aggregates dual-process outputs into a group-level estimate via population-aligned sampling. The second challenge is mechanism duality. Group behavior simultaneously reflects interpretable preference structures and implicit statistical regularities [12], and a single paradigm typically excels at only one end of this spectrum. Fitting-based simulators built on representation learning [5, 13, 25] capture implicit population-level regularities but degrade under long-tail sparsity and semantic shifts. Despite their complementary strengths, effectively bridging the reasoning and fitting paradigms remains an unexplored challenge. Naive ensembling falls short because the two operate with fundamentally different inductive biases, and unconstrained fusion can easily introduce drift without a shared mechanism to align them. We propose Policy-Guided Hybrid Simulation (PGHS), a dualprocess framework for robust group-level counterfactual evaluation. PGHS mines transferable decision policies from behavioral trajectories to form an explicit mechanism alignment layer, grounding both branches with domain-specific empirical logic [15, 17]. This layer anchors two complementary branches inspired by dual-process cognition [12]. A semantic reasoning branch uses policies to constrain LLM inference and prevent over-rationalization. A statistical fitting branch uses policies as conditional priors to absorb implicit regularities. Predictions from both branches are fused under the shared policy space for complementary correction. Our contributions are threefold. • We identify information incompleteness and mechanism duality as two structural barriers to trustworthy group-level user simulation for merchant diagnosis. • We propose mining decision policies from behavioral trajectories as a shared alignment layer to jointly guide LLM reasoning and statistical fitting. • We deploy PGHS on Meituan with 101 merchants and 26k+ trajectories, achieving 8.80% group simulation error with
45.8% and 40.9% improvements over the best reasoning-based and fitting-based baselines.
2
Method
Given a merchant scene 𝑆, an intervention 𝑎 (e.g., a modified price tier or promotional copy), and a user 𝑢, let 𝑦 (𝑢, 𝑆, 𝑎) ∈ {0, 1} denote the binary purchase outcome. Our goal is to predict the group-level purchase rate E𝑢 [𝑦(𝑢, 𝑆, 𝑎)]. As shown in Figure 1, PGHS operates in three phases: (A) policy mining, (B) dual-branch simulation, and (C) group-level fusion.
2.1
Decision Policy Abstraction
This phase distills noisy online behavioral logs into a concise set of transferable decision policies that serve as explicit proxies for the latent mechanism 𝑢 mech . Trajectory Selection. We first construct a high-intent dataset Dchoice = prof 𝑁 by filtering raw interaction data to retain only {(𝑢𝑖 , 𝑆𝑖 , 𝑦𝑖 )}𝑖=1 sessions where a user actively compares multiple merchants before making a definitive purchase or exit. Hierarchical Policy Discovery. We employ a coarse-to-fine pipeline. At the macro level, we serialize each user profile 𝑢 prof into natural language and map it to a dense embedding hprof = Ψenc (𝑢 prof ). K-Means clustering [16] on these embeddings decomposes the population into 𝐾 persona groups P = {𝑃1, . . . , 𝑃𝐾 }. Within each persona 𝑃𝑘 , we leverage an LLM to generate a textual rationale 𝑟 for each high-intent choice instance (𝑢, 𝑆, 𝑦): 𝑟 = MLLM (Explain 𝑦 | 𝑆, 𝑢 prof ).
(1)
Meituan Merchant Business Diagnosis via Policy-Guided Dual-Process User Simulation
We then embed all rationales into a semantic vector space v𝑖rationale = Ψenc (𝑟𝑖 ) and apply HDBSCAN [4] for micro-level refinement: {𝑅𝑘,1, . . . , 𝑅𝑘,𝐽𝑘 }, N𝑘 = HDBSCAN({v𝑖rationale | 𝑖 ∈ 𝑃𝑘 }),
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia.