What is a behavioral foundation model?
A behavioral foundation model is a large model trained on records of behavior, under a general objective, so that one model can be reused for many tasks without being rebuilt for each. Three fields use the idea and mean different things by behavior: how a robot or character moves, what users do in a product, and what customers pay for.
Also called behavior foundation model, large behavior model, large behavioral model, BFM.
The idea in one line
A language model is trained on text and then reused for many jobs. A behavioral foundation model does the same with sequences of actions. It is trained once on a large log of what was done, with no task-specific labels, and the same model is then pointed at whatever task comes up.
What differs between the fields that build behavioral foundation models is whose behavior is in the training log, and that changes what the model can be used for.
In robotics and reinforcement learning
The exact phrase behavioral foundation model is most established in robotics and reinforcement learning. Researchers at Meta use behavioral foundation models to mean agents trained without task rewards that can then be prompted, with no further training, to track a motion, reach a pose or maximize a reward. Their humanoid model, Meta Motivo, was trained with unlabeled motion capture data and was published at ICLR 2025.
Toyota Research Institute uses a related term, large behavior models, for robot policies trained on many manipulation tasks at once. A 2025 study by Toyota Research Institute evaluated large behavior models in simulation and in 1,800 real-world trials, and in August 2025 Boston Dynamics and Toyota Research Institute showed a large behavior model controlling the Atlas humanoid.
In robotics, the behavior in a behavioral foundation model is the movement of a machine or a simulated body. A robotics behavioral foundation model is not a model of what people choose.
In recommendation
In recommendation, the behavior a foundation model learns from is what users did: what they watched, clicked and bought, in order. Meta's generative recommender work trained models with up to 1.5 trillion parameters on users' action sequences and found that quality improved as a power law with compute, the same kind of scaling that made language models work.
Kuaishou's OpenOneRec, released at the end of 2025, co-trained foundation models on general knowledge and 96 million interactions from 160,000 users. Recommendation papers usually say generative recommender or recommendation foundation model in place of behavioral foundation model, and the idea is the same.
In payments and banking
In payments and banking, the behavior a foundation model learns from is a stream of transactions. Stripe announced a Payments Foundation Model in May 2025, which the company says was trained with self-supervision on tens of billions of transactions.
Nubank describes foundation models that read a customer's transactions in time order and represent their financial behavior as an embedding. Nubank's models learn by predicting a customer's future transactions, and Nubank says the models are later fine-tuned for tasks like credit scoring, fraud detection and product recommendation.
What the three share, and what they leave out
Behavioral foundation models in robotics, recommendation and payments all train on what was done, not on what was said about it, and all reuse one model across tasks. Only the recommendation and payments models are about people, and each sees a single slice of a person: their feed, or their card.
None of these behavioral foundation models sees what people say or why they acted. A model of people across settings, trained on what people say and what people do, is called a foundation model of human behavior.
Behavioral foundation model and related terms
- Foundation model of human behavior
the version that is about people specifically, across many settings and tasks, with a representation of each person.
- Foundation model of human cognition
trained on choices people made in psychology experiments, to predict how people think and decide.
- User simulator
a use, not a kind of model: something that plays the user so a product or an agent can be tested.
- Human simulation
the wider practice of producing what people would say or do, which a behavioral model of people can power.
Questions people ask
- What is a behavioral foundation model?
- A behavioral foundation model is a large model trained on records of behavior under a general objective and reused for many tasks. The behavior can be the movement of a robot, the actions of users in a product, or the transactions of customers, depending on the field.
- What is the difference between a behavioral foundation model and a large language model?
- A large language model is trained on text. A behavioral foundation model is trained on sequences of actions, such as movements, clicks or transactions. Some systems combine the two by training a language model further on behavioral data.
- Is a large behavior model the same as a behavioral foundation model?
- Mostly. Large behavior model is the term Toyota Research Institute uses for robot policies trained on many manipulation tasks at once. A large behavior model is the same idea as a behavioral foundation model, applied to robots.
- Which companies build behavioral foundation models?
- Several companies build behavioral foundation models, in different fields. In robotics, Meta has published Meta Motivo and Toyota Research Institute works on large behavior models. In recommendation and payments, Meta, Kuaishou, Stripe and Nubank have described foundation models trained on user actions or transactions. Jean Technologies builds Alyosha, a foundation model of human behavior trained on what people say and do.
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.
Sources
- Tirinzoni et al. (2025), Zero-Shot Whole-Body Humanoid Control via Behavioral Foundation Models, ICLR
- TRI LBM Team (2025), A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation
- Boston Dynamics (2025), Large Behavior Models and Atlas Find New Footing
- Zhai et al. (2024), Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations, ICML
- Zhou et al. (2025), OpenOneRec Technical Report, Kuaishou
- Stripe (2025), Stripe accelerates the utility of AI and stablecoins with major launches
- Stripe, AI in payments
- Nubank (2025), Understanding our customers' finances through foundation models
- Nubank, How Nubank uses transformers to model financial habits at scale