Decomposing the Crowd for Targeted Role-Play: The PRISM Framework for Product Optimization

ACL ARR 2026 January Submission7364 Authors

Keywords: Large Language Models, Generative Agents, Market Simulation, Consumer Profiling, Orthogonal Decomposition, Diagnostic Evaluation, Computational Economics

Abstract: Product design success in consumer-facing markets depends on understanding user needs, yet traditional market research faces high costs and captures only explicit, surface-level requirements. Large Language Models (LLMs) offer scalable market simulation, but existing approaches lack statistical rigor and suffer from mode collapse. We introduce PRISM (Partitioned Role-play in Intelligent Simulated Markets), a framework that bridges unstructured LLM outputs and rigorous economic modeling by decomposing consumer profiles into orthogonal vectors aligned with microeconomic theory. PRISM establishes a theory-grounded diagnostic protocol evaluating dimensional independence and pricing rationality without large-scale ground truth data. We demonstrate that simulated feedback exhibits sufficient structural validity for iterative product optimization, showing clear improvements over baseline methods.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: LLM agents, multi-agent systems, agent evaluation, autonomous agents

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 7364