Test conditions and model performance. (a) Illustration of conditions. The model input comprises a task instruction and a procedure text. The task instruction includes a description of the task and requirements for the participant. The procedure text is a natural-language description of human behavior during the task. (b) Difference in log-likelihood relative to domain-specific cognitive models. The top two rows show the Centaur and Llama models used in Binz et al. [1]. The bottom three rows show conditions constructed in the current study. Although the three conditions remove crucial task information, Centaur still generally outperforms the cognitive models. The results of cognitive models are from Binz et al. [1]. *p < 0.05, **p < 0.01, ***p < 0.001, unpaired two-sided bootstrap, false discovery rate (FDR) corrected.