The foundation model for behavior
One model, every prediction
BehaviorGPT embeds sequences of user actions (purchases, searches, clicks) in a shared semantic space, the way a language model embeds words. Each action is read in the context of the ones before it, so the model learns who a user is from what they do.
Predicting the next action well forces one model to understand the whole user. That understanding then answers the questions you would otherwise build a separate model for. From the same behavior stream, one pretrained model recommends the next action, scores fraud risk, and forecasts churn and demand, with no hand-built features for any of them.
In early research release. Read the technical note.
Inside retail, one model powers search, recommendations, and forecasting — every prediction from the same behavior stream.
What the model does
One pretrained model does all four, learned from behavior alone, with no task-specific features or labels.
embed
Represent agents and items as embeddings from behavior alone.
predict
Predict the next action in a session from its trajectory.
transfer
Reuse one pretrained model across recommendations, search, and fraud scoring.
score
Score intent and risk in real time as a session unfolds.
FAQs
What is BehaviorGPT?
BehaviorGPT is a model that represents user behavior (purchases, searches, clicks, and other actions) as embeddings in a shared semantic space, much as a language model embeds words and sentences. It learns these representations directly from sequences of behavior, without hand-engineered features.
How is this different from a normal recommendation engine?
A typical recommender scores items against a user profile or matches similar users. BehaviorGPT instead models the trajectory of a session: the actions leading up to a moment reshape the meaning of the next one. That context lets it disambiguate intent rather than treat each action in isolation.
Can BehaviorGPT be used for fraud detection?
Yes. Fraud is a trajectory problem: the same action can be routine or suspicious depending on the sequence that led to it. Because BehaviorGPT models that sequence directly, it can score risk in real time as a session unfolds and surface rare, high-value patterns that fixed rules and hand-engineered features tend to miss. The same pretrained model that powers recommendations can be adapted to fraud and risk scoring.
What kind of data does it use?
It learns from chronological sequences of actions (transactions, searches, sessions, and interactions) treated as event streams. It does not require labels or manually engineered features; the order of events is the signal.
Is this available as an API or product yet?
Not yet. BehaviorGPT is in early research release. Joining the waitlist registers your interest for early access as we open capacity.
When will I get access after joining the waitlist?
We're onboarding participants gradually. After you join, we'll reach out by email as access opens for your use case. There is no fixed date yet.
Will my data be used to train the model?
Waitlist signups are used only to contact you about access. Any future data use would be governed by an explicit agreement. Nothing you share now is used for training.
Who is this for?
BehaviorGPT is aimed at teams working on fraud and risk, trust and safety, e-commerce personalization, search and recommendations, and credit and lifetime-value modeling, as well as ML research teams studying representations of behavior.
Is there a technical paper or documentation available?
A technical note is in preparation. We'll share it with waitlist members, and you can follow our ongoing work at research.unboxai.com.