A foundation model to predict and capture human cognition

https://www.nature.com/articles/s41586-025-09215-4

Centaur: Llama 3.1 70B, frozen 4-bit base, rank-8 QLoRA adapters (about 0.15% of parameters), one epoch, about five days on one A100 80GB. Psych-101: 160 experiments, 60,092 participants, 10,681,650 choices. Beats 14 cognitive models on held-out participants; generalizes to new cover stories, structures, domains. Limits: WEIRD bias, mostly learning and decision tasks, 8B Minitaur weaker out of distribution.

LLM models of individual human choice