Large Behavior Model: language model learned to predict purchases as digital twin

Researchers in July 2026 presented Large Behavior Model (LBM) on arXiv — a language model trained directly on large-scale retail purchase data and capable of serving as a "promptable digital twin" of a customer, predicting their purchase decisions.

How the model is structured

LBM describes customer behavior through a single "Person-Environment" formulation: the customer's state is defined by a behavioral profile built on their past purchase history, and product context is added through retrieval-augmented generation (RAG) — the model "pulls" relevant product information during generation rather than storing it entirely in its weights.

This approach differentiates LBM from classical recommendation systems: instead of treating customers 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 same way a language model interprets dialogue.

Model training proceeds in three stages:

• Continued pre-training — fine-tuning a base language model on "verbalized" (converted to text format) behavioral purchase data
• Supervised fine-tuning — precise adjustment for generating specific purchase decisions
• Reinforcement learning with verifiable rewards — used to calibrate the model based on 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, allowing judgment of its generality within the domain. The authors evaluated LBM on five types of tasks: purchase prediction, hard-negative discrimination, basket completion, response to promotional offers, and cross-domain voucher redemption when transferred between different retailers.

Across all these tasks, LBM consistently outperformed frontier-level universal language models, while demonstrating strong knowledge transfer in zero-shot mode — that is, without additional fine-tuning — and after fine-tuning on new retailers and decision types the model was not originally trained on. This matters for practical application: retailers don't need to collect and label massive datasets from scratch for each new task or store network.

What provided the greatest effect

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

What this means

The work demonstrates that behavioral data from transaction histories can be effectively encoded into a language model and yield a scalable foundation for "digital twins" of customers — systems applicable immediately to multiple retail analytics tasks: from recommendations and marketing to simulating user reactions to promotional offers without the need to conduct expensive A/B tests on real customers.

Particularly valuable from a practical standpoint is that the model is called "promptable": an analyst can address the digital twin of a customer in natural language and get a prediction or explanation of a decision, rather than only a numerical probability score. This closes the gap between the data science team that builds the model and business users — marketers and category managers — who need to quickly validate hypotheses about customer behavior without writing separate analytical code for each question.