Tag
This paper investigates whether small foundation models fine-tuned on human behavioral data can serve as cognitive proxies, finding that scale matters little in-distribution but larger models generalize better out-of-distribution.
This paper trains 14 small language models (135M to 14B parameters) on Psych-101, a dataset of 10.7 million trial-level human choices, finding that small models suffice for in-distribution matching while larger models generalize better out-of-distribution. Diagnostics show that masking stimuli and feedback destroys most learned information, indicating choice history alone is insufficient.