How humans build and maintain abstract knowledge from sparse, noisy streams of experience is a longstanding challenge in cognitive science. Any computational account must satisfy three desiderata: (1) data and compute efficiency; (2) representation of graded uncertainty to enable intelligent inquiry and information gathering; (3) flexibility to mentally represent the endless range of concepts people can learn. We introduce a model that meets these criteria by encoding symbolic knowledge as "mental programs"—a blend of natural language and source code—and performing sequential inference with LLM‑guided Bayesian learning algorithms. Across diverse behavioral studies the model reproduces quantitative signatures of human inductive learning and active inquiry, such as anchoring, garden‑pathing, and related effects. In contrast, pure LLMs or classic Bayesian models either fail on the core task, do not capture human behavior, or require prohibitive computational cost. These findings suggest that humans may continuously grow knowledge by maintaining many hypotheses spanning language‑like and program‑like representations, while a bottom‑up neural mechanism (an LLM) makes inference tractable and learnable.
Review