Large Behavior Model: arXiv paper on a retail digital twin

Large Behavior Model, submitted to arXiv on 8 Jul 2026, is a promptable digital twin that learns customer decision making directly from large-scale retail transactions. The paper, by Wachiravit Modecrua, Krittin Pachtrachai and Touchapon Kraisingkorn (arXiv:2607.06993), describes training and evaluation across multiple retail decision tasks.

What is the Large Behavior Model?

The Large Behavior Model (LBM) is a language-model-based digital twin that represents customer state with a behavioral profile derived from historical purchases and incorporates product context via retrieval-augmented generation. The authors frame the problem as a unified Person-Environment formulation so that customer state and product context are explicit inputs to the model, enabling promptable decision generation for retail scenarios.

LBM is presented as a single system that can be prompted to make retail decisions, simulate behavior, and transfer across retailers and decision domains. The arXiv submission spans 17 pages and includes 5 figures documenting the approach and results.

How was the model trained and evaluated?

The model was trained in three stages: continued pre-training on verbalized behavioral data, supervised fine-tuning for decision generation, and reinforcement learning with verifiable rewards to calibrate evidence-based outputs. Retrieval-augmented generation supplies product context, and the authors apply retrieval during both training and inference to improve performance.

Evaluation covered purchase prediction, hard-negative discrimination, basket completion, promotion response, and cross-domain voucher redemption. The paper states that LBM consistently outperforms frontier general-purpose language models on in-domain retail tasks and shows strong zero-shot and fine-tuned transfer across retailers and decision domains. Ablation studies in the paper identify continued pre-training as the primary driver of behavioral generalization, retrieval as most effective when used at both training and inference time, and reinforcement learning as improving the model's reliance on explicit behavioral evidence rather than generic language-model priors.

Why it matters

LBM moves behavior modeling away from purely predictive or purely simulated approaches by grounding decisions in transaction histories and by making customer state explicit and promptable. If continued pre-training on verbalized purchase histories and retrieval-augmented context reliably yields better in-domain performance and cross-domain transfer, retailers could adopt a single, adaptable model for prediction, personalization and simulation rather than maintaining many specialized systems. That shift affects recommendation, marketing, promotion planning and decision support workflows that depend on accurate, explainable behavior models.

What to watch

Look for code or data releases linked to arXiv:2607.06993 and for replication of the paper's ablation claims: specifically, whether continued pre-training remains the largest contributor to behavioral generalization, and whether retrieval applied at both training and inference yields the same gains in independent evaluations.