A new hybrid system uses AI agents to fix human-built models of how people reason, instead of starting from scratch.
Researchers built a pipeline that takes existing probabilistic programs, hand-coded models of human cognitive algorithms, and hands them to a system of LLM agents. The agents find mismatches between a model's predictions and real human behavior, propose specific code-level edits within constraints set by the researchers, and check that those edits do not break the model's structure. A probabilistic inference module then recomputes how well the revised model explains the behavioral data. The team tested this on a problem-solving task designed to surface different reasoning strategies, and the revised models consistently fit human behavior better than the originals.
This sits between two approaches that each have obvious flaws: hand-built cognitive models are interpretable but slow to improve, while pure LLM-generated models are scalable but opaque and mostly untested on anything as complex as recovering an algorithm from behavior. By having AI agents make targeted edits instead of writing a model from a blank slate, researchers keep the parts of their theory they trust while letting a machine surface fixes they missed. The paper also reports that a small, recurring set of code tweaks accounted for most of the improvement, hinting that human reasoning strategies on this task cluster into a manageable number of variants rather than infinite idiosyncrasy.
It is a narrow, technical result, but it points to a more useful role for LLMs in science: patching existing theories instead of replacing them outright.