Can matching human actions prove an LLM simulation explains behavior?

The paper argues that LLM social simulations are "primarily evaluated for behavioral fit," meaning whether agents reproduce the actions or response distributions of the people they stand in for. That target is adequate when a simulation only predicts behavior. It is not adequate when a simulation is asked to "explain human behavior, diagnose barriers, and compare large-scale interventions," because those uses "depend on understanding why people acted a certain way, not just what they did." The abstract's reason is that "behavior underdetermines the reasoning process behind it": staying silent may reflect disinterest or suppressed speech, and an unanswered call may reflect distrust of the caller or limited phone access.

The proposed replacement is "representational adequacy." It uses LLM reasoning traces to ask whether a simulation's scenario–reasoning–action triples preserve the reasoning behind the behavior, faithfully to the population and scenarios being simulated. The introduction supplies the enabling observation. Classical agent-based models make the researcher write decision rules, whereas an LLM agent "generates its own behavior and can be asked what it weighed before acting." The trace can be produced before the action, stored as structured state, elicited afterward, or reconstructed from logs, and in every case it is treated as "a hypothesis about the process behind the agent's action" and not as a record of it. The conclusion frames the target as preserving "what people considered, feared, and almost did."

This sits close to Can we measure reasoning quality beyond output plausibility?, which also moves evaluation from outputs to the structure of belief formation. The difference visible in the excerpt is scope: this paper names the use cases, explanation, diagnosis and intervention comparison, that make behavioral fit insufficient, and it frames the target at the level of the simulated population. It also bears on Can AI agents learn people better from interviews than surveys?. A response-replication score of that kind is a behavioral-fit result, and on this paper's argument it would not by itself license claims about why the simulated people answered as they did. Can structural causal models automate social science with language models? reports that simulated effects get their signs right but not their magnitudes, which is the intervention-comparison use this paper says fit alone cannot support. How do we generate realistic personas at population scale? addresses outcome distributions across a population. The present paper adds a second axis: whether the reasoning behind those outcomes is also represented.

The excerpt does not say how representational adequacy is scored, what reference stands in for the population's actual reasoning, or whether any simulation was tested against it. It reports no results. It also leaves open how far a reasoning trace tracks the process that produced the action, which the authors themselves call a hypothesis. The claim that survives at this evidence level is narrow. Behavioral fit cannot tell apart different reasons behind the same action, so it cannot support explanatory claims alone. Whether traces can fill that gap is a proposal here, not a demonstrated result.