Tag
UC Berkeley researchers propose a hybrid system treating cognitive algorithm discovery as a program refinement problem, where LLM agents revise human-authored probabilistic programs based on behavioral data mismatches, consistently improving model fit in a problem-solving paradigm.