Unbox AI builds foundation models that learn from what people do, not what they say. Our Large Behavioral Models (LBMs) are trained on chronological sequences of human actions—purchases, transactions, interactions—across billions of users. This research blog documents our work advancing this new paradigm.
A 12.5B-parameter Large Behavioral Model pretrained on 150 billion user actions (3.6 trillion tokens) across retail, engagement, and payments. With no target-specific training, it outperforms specialists trained on held-out public datasets, leads on 30 of 31 dataset–task pairs zero-shot, and ranks a full catalog in under a millisecond per query.
Our research has roots in topology—the mathematical study of shape and structure—which provides rigorous tools for unsupervised and self-supervised learning. Because topological methods generalize naturally to discrete structures like graphs, this perspective led us to graph learning and the insight that Transformers can be viewed as Graph Neural Networks operating on sequential (line-graph) structures. This unifying view enables us to extend language-modeling concepts beyond text to new domains, such as behavioral sequences.