Large Language Models Predict Human Social Behavior via Interpretable Mechanisms

Introduces MindEvolve, an autonomous workflow that uses LLMs to generate interpretable symbolic models of cognition for predicting behavior in social interactions, rather than only imitating behavior. Prior work emphasized behavioral imitation with limited attention to transparent or interpretable models of the cognitive mechanisms underlying human decisions; this work shifts toward generating interpretable symbolic models. The paper proposes and systematically evaluates this workflow as a methodological contribution; implementation details are not expanded in the provided summary-level text.

Systematically evaluates the modeling capabilities of multiple LLMs across a battery of socioeconomic games covering four core domains of social cognition: economic preferences, social preferences, social reasoning (theory of mind), and recursive planning. Extends evaluation across four core domains of social cognition rather than a single game or a single preference dimension. The evaluation is based on a battery of socioeconomic games spanning four domains, constituting a multi-task, multi-model systematic assessment design.

Most LLMs robustly capture economic and social preferences in relatively simple strategic settings, and in some cases generate novel models that integrate broader knowledge than those proposed by human experts. Suggests LLMs may go beyond reproducing existing preference structures to proposing novel symbolic models with broader knowledge integration. Conclusions come from evaluating multiple models, with expert human evaluators assessing interpretability and theoretical coherence; the number of models and evaluators is not given in the provided text.

Capacity to model more complex psychological processes remains limited, but a subset of state-of-the-art models demonstrates promising performance in capturing higher-order reasoning processes such as theory of mind and recursive planning. Distinguishes simple preference modeling from higher-order reasoning modeling, identifying the latter as both a current boundary and a potential breakthrough point. This judgment is based on comparisons across cognitive domains within the same battery of games, a layered capability description rather than a single-metric conclusion.

The work targets behavior prediction in social interactions and applies to social cognition research settings built on socioeconomic games; its value lies in offering researchers interpretable symbolic cognitive models and an expert-evaluation pathway. It is most relevant to researchers and theory builders who want to use LLMs to construct testable cognitive hypotheses and who care about model interpretability.

The provided text is summary-level and does not include the specific model list, game task details, the number of expert evaluators or scoring method, nor quantitative comparisons across domains; therefore the exact magnitude of capability differences between cognitive domains and the specific scope of the 'subset of state-of-the-art models' still require consulting the original figures and experimental details to confirm.