Agent psychonomicus: Psychological mechanisms emerge in cognitive agents simulated by large language models

Researchers increasingly use large language models (LLMs) to simulate participants in psychological experiments. Yet researchers observe only the LLM's answer and cannot tell whether a manipulation acted through the psychological mechanism that a theory proposes. We therefore introduce Agent psychonomicus, a class of cognitive models for testing each step of the mechanism proposed by a process theory. We implement it as the Minimal Viable Cognitive Agent (MVCA), in which an ensemble of 12 LLM agents carries out general cognitive processes, such as appraising how a situation bears on the self, and states the results as first-person process statements. Using three different LLMs, we conducted 57,147 cognitive simulations of 60 published experiments and analyzed over six hundred thousand process statements. We found that the theorized psychological mechanism, or the mediator, emerged from general processes simulated by MVCA. Inducing the theorized mechanism directly in process statements had a causal effect on the experiments' outcomes. Among experiments re-examined in large preregistered replications, MVCA's cognitive agents showed a significantly larger share of the original effects that replicated than of those that did not, suggesting that MVCA's cognitive simulations of human participants are diagnostic and address the false-positive results of silicon samples in existing research.