Jean is now available in Claude. See how
Jean Technologies Icon
Jean Technologies Text

Glossary

What is a foundation model of human cognition?

A foundation model of human cognition is a single model meant to predict and simulate how people think and decide across many kinds of task, where earlier cognitive models each explained one. The term comes from Centaur, a model published in Nature in 2025 that was trained on the choices people made in psychology experiments.

Also called cognitive foundation model, foundation model of cognition, Centaur.

Where the term comes from

Psychology has long built one model per phenomenon: a model of memory, a model of risky choice, a model of learning from reward. Marcel Binz, Eric Schulz and their colleagues proposed a single model for all of them. Their preprint in 2024 was titled Centaur: a foundation model of human cognition, and the paper appeared in Nature in 2025 as A foundation model to predict and capture human cognition.

How Centaur was built

Centaur's authors assembled a dataset called Psych-101. Psych-101 holds 160 psychology experiments, written out in plain language, with about 10.7 million choices made by 60,092 participants, one trial at a time.

Centaur is an open language model, Llama 3.1 70B, fine-tuned on that data. The base model stayed frozen and small adapter layers were trained on top of it, a method called QLoRA. So Centaur starts with what a language model knows about the world and learns from the experiments how people actually choose.

What it showed

Centaur captured the behavior of participants it had never seen better than existing cognitive models did. It also generalized: to new cover stories for a known task, to changed task structures, and to entirely new domains.

Centaur's authors also report that after fine-tuning, Centaur's internal representations lined up more closely with human neural activity.

What it is not

Centaur's training data is laboratory behavior: choices in controlled tasks with defined options. Psych-101 is not everyday behavior such as what people buy, watch, or say in an interview, and Psych-101 does not follow any one person across their life.

A foundation model of human cognition differs from a foundation model of human behavior, which is defined by training on records of what people did in the world and by learning a representation of each person. The two ideas are close relatives. Centaur is strong evidence that behavioral data teaches a model things that text alone does not.

Why it matters outside psychology

Anyone who simulates people, for research or for product testing, depends on a model that chooses the way people choose. General assistants are trained to be helpful, which is a different goal. Centaur shows that training on recorded human choices moves a model toward people, and that the gain carries to tasks outside the training set.

Foundation model of human cognition and related terms

Questions people ask

What is a foundation model of human cognition?
A foundation model of human cognition is a single model meant to predict and simulate how people think and decide across many kinds of task. The term comes from Centaur, a model published in Nature in 2025 by Marcel Binz, Eric Schulz and colleagues.
What is Centaur and who built it?
Centaur is a foundation model of human cognition from Marcel Binz, Eric Schulz and colleagues, published in Nature in 2025. Centaur is the language model Llama 3.1 70B fine-tuned on Psych-101, a dataset of choices people made in psychology experiments.
What is the Psych-101 dataset?
Psych-101 is the dataset Centaur was trained on. Psych-101 contains 160 psychology experiments written in plain language, with about 10.7 million choices from 60,092 participants.
What is the difference between a foundation model of human cognition and a foundation model of human behavior?
A foundation model of human cognition, such as Centaur, is trained on choices made in laboratory experiments. A foundation model of human behavior is trained on records of what people did in the world and learns a representation of each person. The two are close relatives.
Which companies build foundation models of human behavior?
Centaur is an academic model, not a company's product. Among companies, Jean Technologies builds Alyosha, a foundation model of human behavior, and Diachronic describes its work as foundation models to predict people. Meta, Kuaishou, Stripe and Nubank have described large models trained on behavior inside one product, such as a feed or a payments network.

What Jean offers

Jean Technologies is the #1 AI market research platform to learn from real people and digital twins.

  • Jean Technologies offers an AI interviewer that runs AI-moderated interviews with real people, by voice or text, in any language, for research teams that need the reasons behind a number from more people than they can interview themselves.
  • Jean Technologies offers digital twins, synthetic respondents each built from a real person's interview, for teams doing AI market research who want to test concepts, prices, bundles and messages before fieldwork, a practice also called decision simulation.
  • Jean Technologies offers simulated users that use a website, app, chatbot or voice agent before launch, for product teams that want to see where people drop off and why.
  • Jean Technologies builds Alyosha, a foundation model of human behavior trained on what people say, what they do, and why, for companies that want digital twins and simulations run on a model of people. Alyosha has the highest overall score on Jean's open evaluations of how well models simulate people: 84.6 across the SOUL benchmark tasks.