AutoCog: AI agent system discovers psychological theories autonomously
LLM-driven 'cognitive scientist' beats humans at building decision-making theories
AutoCog, a fully autonomous agentic-AI system, closes the loop in psychological theory discovery. Large-language-model agents advocate competing theories as executable cognitive models, design experiments, recruit online participants, score theories based on generative performance, diagnose failures, and synthesize better successors. In decision-making experiments with simulated behavior, it recovered known strategies. With human participants, it produced theories that outperformed the established theories it was seeded with and generalized to held-out studies.
- AutoCog uses LLM agents to compete, design experiments, collect human data, and iteratively improve cognitive theories
- Discovered a novel theory of diminishing sensitivity to feature values in multi-cue decision-making, confirmed via preregistered study
- Outperformed established seeded theories and generalized across two different experimental settings
Why It Matters
Automates the most creative bottleneck in cognitive science—theory building—potentially accelerating discovery of human decision-making models.