Proceedings of the Annual Meeting of the Cognitive Science Society

A central aim of human-computer interaction is to build systems that anticipate users' goals by modeling their underlying cognitive processes. Recent work on general user models infers user states from behavioral observations such as screenshots and action logs. Behavior alone, however, underdetermines the algorithmic-level processes (Marr, 1982) that generate it: distinct reasoning trajectories routinely yield identical actions, leaving inferred user states confounded with the inductive biases of the inferring model. We argue that the think-aloud method (Newell & Simon, 1972; Ericsson & Simon, 1993) offers process-level evidence that behavioral traces cannot, and that recent advances in automated transcription and LLM-based coding make it feasible at ecological scale. Reframing user modeling as the recovery of structured reasoning traces, rather than the prediction of behavior, positions cognitive science to contribute foundational representations to a problem currently dominated by behavioral proxies.