Small Foundation Models of Human Cognition and Behaviour

Large language models finetuned on human behavioural data have emerged as general-purpose cognitive proxies, but whether these models process task structure or exploit statistical shortcuts in choice sequences, remain open questions. We train eight models from 0.6B to 14B parameters across two architecture families on Psych-101, a dataset of 10.7 million trial-level choices from 160 experiments. Performance on held-out participants plateaus by 4B parameters, with a 0.6B model matching a reproduced 70B baseline; out-of-distribution evaluation on 15 unseen experiments confirms transferable gains with a steeper scaling gradient. To investigate what information these models rely on, we decompose each prompt into four channels -- task instructions, experimental stimuli, outcome feedback, and choice history -- and systematically remove each across 32 experiments. Removing stimuli and feedback destroys 75.7% of learned information and pushes models below chance, demonstrating that choice history alone does not account for performance. Order permutation tests further reveal invariance on tasks with independent trials but sensitivity where trial order is determined by prior responses. These results establish that small-scale cognitive finetuning produces task-adaptive proxies whose predictions reflect experimental structure rather than surface-level prompt statistics.