A model trained on what people actually do, that learns a representation of each person and serves many tasks at once: predicting what someone will do next, simulating how a population would respond, ranking what to show them.
Where the term comes from
Stanford researchers coined foundation model in 2021: a model trained on broad data, generally with self-supervision at scale, that can be adapted to a wide range of downstream tasks. Large language models are the best-known example. A foundation model of human behavior applies the same idea to people rather than text.
The definition
Several fields have each proposed their own version: a foundation model of human cognition, a general user model, a large behavioral model. Our working paper rolls them into one checkable definition. A foundation model of human behavior is a model that:
Generality is the test: a good person-level representation should carry predictive power across many tasks, including ones it was never trained for. More general only means something once the set of tasks is fixed, which is what a benchmark does.
Three fields are converging on it
Three fields predict, simulate and model people for different reasons, and each has been fitting the same objective without saying so: predict a person's next action from their history, each with the history cut down to the part it can see.
Are large language models already one?
Partly. Text is written by people, so a model of text is a model of human behavior, and language models do reproduce some behavioral experiments. But they were not built for the role. Post-training collapses them into a single helpful assistant, which narrows the range of human answers pre-training absorbed, and their training never saw any one person's history.
What is still missing
Deep learning reached goals like this in other domains once three things were in place: a generic objective, a model that scales under it, and a corpus large enough to scale on. For people, the first two are in place. The third, a corpus of what people do that spans the silos holding it, is not. Assembling it, with consent, is the open step.
How it is measured
A model of people is graded against people: held-out behavior, real survey answers, and how consistent people are with themselves. We collect the published benchmarks and run our own on the open evaluations page.
Read further
The full argument is in the working paper, Toward a Foundation Model of Human Behavior, and the short version in the essay Teaching Machines to Understand Humans. Our own model is Alyosha.
