Automated Discovery of Sparse and Interpretable Cognitive Equations

Discovering computational models that explain human cognition and behavior remains a central goal of cognitive science, yet the reliance on hand-crafted equations limits the range of cognitive mechanisms that can be uncovered. We introduce SPICE (Sparse, Interpretable Cognitive Equations), a framework that automates the discovery of mechanistically interpretable cognitive models directly from behavioral data. SPICE fits recurrent neural networks to capture latent cognitive dynamics and then applies sparse equation discovery to extract concise mathematical expressions describing those dynamics. Theory-guided priors make the approach data- and compute-efficient, while a hierarchical design reveals individual differences in the algorithmic structure of cognitive dynamics rather than in parameters alone. In simulations, SPICE accurately recovered the structure and parameters of known reinforcement learning models. Applied to human behavior in a two-armed bandit task, it uncovered new equations that outperformed existing models and revealed structural alterations in reinforcement learning mechanisms among participants with depression, such as a loss of nonlinear exploration dynamics regulating behavioral flexibility. This approach provides systematic insights into structural individual differences in cognitive mechanisms and establishes a foundation for automated discovery of interpretable behavioral models.