New 4-Level Scale Evaluates LLM-Based Agent Models for True Explanation

Large language models now power agent-based models (ABMs) for simulating human behavior on social platforms or in game theory. But a key problem, according to a new paper accepted at ACM FAccT 2026, is conflating predictive capability with genuine explanation. Authors Zhao, Pham, and Vincent draw on philosophy of science to propose a formal scale—the Mechanism Plausibility Scale—that separates a model's ability to reproduce a phenomenon (generative sufficiency) from its ability to show how that phenomenon is produced by organized entities and activities (mechanistic plausibility).

The scale has four levels, from simple replication to full mechanistic insight. This framework gives modelers a shared language to judge whether their LLM-based ABMs are just pattern-matchers or true explanatory tools. The paper also clarifies the distinct roles of predictive versus explanatory models, helping researchers avoid overclaiming insight from simulations that merely mimic outcomes. For practitioners building AI-powered social simulations, this means more rigorous evaluation standards—and better understanding of when an LLM-based agent is actually causal versus just statistically fitting data.