Large Behavior Model Creates Promptable Retail Customer Digital Twin

Traditional customer behavior modeling often sacrifices explainability for predictive accuracy or lacks grounding in real data. A new approach, the Large Behavioral Model (LBM), aims to bridge this gap by learning customer decision-making directly from extensive retail transaction histories. This model uses a Person-Environment framework, representing customer state via a behavioral profile from past purchases and incorporating product context through retrieval-augmented generation.The LBM is trained through a multi-stage process: continued pre-training on verbalized behavioral data, supervised fine-tuning for decision generation, and reinforcement learning with verifiable rewards for evidence-based calibration. Evaluations across tasks like purchase prediction, basket completion, and promotion response show the LBM consistently outperforms leading general-purpose language models on in-domain retail tasks. It also exhibits strong zero-shot and fine-tuned transfer capabilities across different retailers and decision domains.Ablation studies reveal that continued pre-training is crucial for behavioral generalization, while retrieval is most effective when used during both training and inference. Reinforcement learning enhances the model's reliance on explicit behavioral evidence over generic language model priors. These findings confirm that behavioral knowledge from transaction histories can be effectively learned by language models, providing a scalable foundation for customer digital twins and advanced behavior simulation.