Large Behavior Model: A Language Model That Predicts Purchases Like a Digital Twin

In July 2026, researchers presented Large Behavior Model (LBM) on arXiv. It is a language model trained directly on large-scale retail purchase data, and it can act as a "promptable digital twin" of a shopper, predicting their purchase decisions.

How the model works

LBM describes customer behavior through a single "person-environment" formulation (Person-Environment). The shopper's state is defined by a behavioral profile built from their past purchase history, while product context is added through retrieval-augmented generation (RAG): the model pulls in relevant product information while generating its answer, rather than storing all of it in its weights.

This sets LBM apart from classic recommender systems. Instead of treating a shopper as a set of features (age, order history, segment), the model processes behavior as a sequence of decisions that can be described in text and interpreted the way a language model interprets a dialogue.

Training takes place in three stages:

• Continued pre-training: further training of a base language model on "verbalized" purchase behavior data (converted to text format)
• Supervised fine-tuning: tuning for generating specific purchase decisions
• Reinforcement learning with verifiable rewards: used to calibrate the model against actual data rather than general language preferences

What tasks the model was tested on

The authors tested the model on a broad set of typical retail tasks rather than a single narrow metric, which allows a judgment of its versatility within the domain. They evaluated LBM on five task types: purchase prediction, hard-negative discrimination, basket completion, response to promotions, and cross-domain voucher redemption.

Across all these tasks, LBM consistently outperformed frontier-level general-purpose language models. It also showed strong knowledge transfer in zero-shot mode (that is, without additional fine-tuning) and after fine-tuning on new retailers and decision types it was not originally trained on. This matters in practice: retailers do not need to collect and label huge datasets from scratch for every new task or store chain.

What had the biggest effect

Ablation experiments showed that continued pre-training, the stage of further training on "verbalized" purchase data, is the main driver of the model's ability to generalize behavior to new situations. Retrieval (pulling in product context) was most effective when applied both during training and at inference, rather than at only one of them. Reinforcement learning helped the model rely more on explicit behavioral evidence in the data than on general language "assumptions" inherited from the original language model.

What this means

The work shows that behavioral data from transaction histories can be effectively encoded into a language model, yielding a scalable foundation for shopper "digital twins": systems applicable to several retail analytics tasks at once, from recommendations and marketing to simulating user responses to promotions without running expensive A/B tests on real customers.

The model's "promptable" nature has particular practical value: an analyst can query the shopper's digital twin in natural language and get a prediction or an explanation of a decision, not just a numerical probability estimate. This narrows the gap between the data science team that builds the model and business users, such as marketers and category managers, who need to test hypotheses about customer behavior quickly without writing separate analytical code for each question.