Large Behavior Model: A New Digital Twin for Retail Customers

The Large Behavior Model (LBM) is a groundbreaking approach to customer behavior modeling, introduced by Wachiravit Modecrua and colleagues on July 8, 2026. This model learns customer decision-making directly from extensive retail transactions, aiming to enhance recommendation systems and marketing strategies.

Understanding the Large Behavior Model

The LBM differentiates itself by integrating customer behavioral profiles derived from historical purchases with contextual product information. It employs a unified Person-Environment formulation, which allows for a more accurate representation of customer states.

Through a process of continued pre-training on verbalized behavioral data and reinforcement learning, the model achieves improved decision generation and evidence-based calibration. These techniques enhance the model's predictive capabilities, making it a powerful tool for retailers aiming to optimize customer engagement.

Performance Evaluation of LBM

The LBM has been rigorously evaluated across several key metrics, including:

• Purchase prediction
• Hard-negative discrimination
• Basket completion
• Promotion response
• Cross-domain voucher redemption

Results indicate that LBM consistently outperforms conventional language models, demonstrating strong capabilities in zero-shot and fine-tuned transfer across various retailers and decision-making domains.

Key Findings and Implications

Ablation studies reveal that the continued pre-training process is crucial for behavioral generalization, while retrieval-augmented generation is most effective during both training and inference stages. Moreover, reinforcement learning enhances the model's reliance on explicit behavioral evidence, shifting away from generic language-model priors.

These findings suggest that leveraging transaction histories can significantly improve the way language models learn and simulate customer behavior, paving the way for scalable customer digital twins.