New AI Model Creates Promptable Digital Twins of Retail Customers
Retail behavior simulated via LBM outperforms GPT-4 and Claude on purchase prediction
Deep Dive
Researchers introduced the Large Behavioral Model (LBM), a promptable digital twin of retail customers trained on transaction data. It uses a Person-Environment formulation with retrieval-augmented generation and reinforcement learning. LBM consistently outperforms frontier general-purpose language models on in-domain retail tasks, enabling zero-shot transfer across retailers for scalable customer simulation.
Key Points
- LBM uses continued pre-training on verbalized transactions as the primary driver of behavioral generalization.
- Retrieval-augmented generation applied during both training and inference significantly boosts performance.
- Reinforcement learning with verifiable rewards improves reliance on explicit behavioral evidence over generic language priors.
Why It Matters
Enables accurate, explainable customer digital twins for personalized marketing and decision support at scale.