Induction and Inquiry Through Probabilistic Reasoning Over Language and Code

Summary

Researchers propose a computational model of human abstract learning that represents hypotheses as programs combining natural language and source code. An LLM-guided Bayesian process sequentially infers and revises these programs, reproducing behavioral patterns such as anchoring and garden-path effects. The approach is more behaviorally faithful or computationally practical than pure LLMs and classical Bayesian models in the reported studies.