The bitter lesson for cognitive science
“You can’t play 20 questions with Nature and win.”
— Allen Newell
So let’s stop asking one question at a time.
Scaling cognitive science through agentic pipelines, richer data, and automated discovery.
Psyche
Younes Strittmatter, Tom Griffiths, Akshay Jagadish
The largest standardized dataset of human behavior
- 242,769 participants
- 37,343,110 behavioral responses
- 550 experiments
Large, carefully curated datasets have transformed science. The Protein Data Bank helped lay the foundation for breakthroughs such as AlphaFold. We believe cognitive science is approaching a similar turning point.
Psyche is an agentic system for building a massive, standardized database of human behavior. It searches the scientific literature for papers with publicly available datasets, extracts the underlying experimental data, and iteratively standardizes and validates them through automated feedback.
But collecting the data is only the beginning.
We are also developing an agentic research pipeline that can query Psyche, connect evidence across hundreds of experiments, and help researchers ask—and answer—questions that were previously impossible to study at scale.
Be among the first to explore Psyche
Get an e-mail notification on realease →
AutoCog
Akshay K. Jagadish*, Younes Strittmatter*, Nori Jacoby, George Kachergis, Eric Schulz, Nathaniel Daw, Suyog H. Chandramouli*, Thomas L. Griffiths*
Autonomous systems are beginning to transform science by closing the loop between theory, experiment, and data. In cognitive science, however, theory generation remains largely manual: researchers must still turn model failures into better explanations. We introduce the Automated Cognitive Scientist (AutoCog), an agentic AI system that automates this process. Competing agents formulate executable cognitive theories, design discriminating experiments, recruit online participants, evaluate models against behavioral data, diagnose their failures, and synthesize improved successors. Repeated cycles enable AutoCog to search jointly over theories, models, and experiments.
In decision-making tasks, AutoCog recovered known strategies from simulated data, including unconventional ones, showing that its discoveries were guided by evidence rather than limited to the priors of its language models. With human participants, it developed theories that outperformed established baselines and generalized to held-out studies across two experimental settings. It also proposed a novel account of multi-cue choice based on diminishing sensitivity to feature values, whose predictions were confirmed in a preregistered experiment. AutoCog turns cognitive theory-building into an explicit, executable, and cumulative process.
