Preprint describes AI behavioral profiles built from retail transactions

What happened

The authors’ revised preprint, “Large Behavior Model: A Promptable Digital Twin of the Retail Customer,” reports a system trained on verbalized behavioral data from retail transactions. It represents each customer through a profile derived from historical purchases and adds product context through retrieval-augmented generation, which retrieves relevant information for a model response.

The reported primary retail cohort contained 159,667 training transactions, 40,665 test transactions, 1,500 customers with held-out baskets, and 38,516 catalog products. Records included a masked customer identifier, timestamp, product identifier, quantity, price, discounts, and final spend.

The authors evaluated purchase prediction, hard-negative discrimination, basket completion, promotion response, and cross-domain voucher redemption. They report stronger results than frontier general-purpose language models on in-domain retail tasks, but the work remains a preprint rather than peer-reviewed or independently replicated research.

Why it matters

The paper offers task-specific evidence that transaction histories can support detailed AI-generated consumer profiles. Separately, FTC staff material says some surveillance-pricing respondents can use granular consumer data for individualized prices and discounts. Neither source shows this research system has been deployed for targeting, pricing, or other commercial use.

What remains unclear

• The accessed paper does not verify the retailer, transaction dates, jurisdiction, data-acquisition terms, customer notice or consent basis, retention practices, or re-identification-risk testing.
• The authors report weaker transfer when useful signals depend on platform-specific identifiers or engineered features. They also did not quantitatively validate open-ended generation of customer reviews, rationales, or survey responses.
• No source in the dossier identifies a retailer using the model or shows that its profiles were linked to named people, used to target consumers, or caused harm.

Sources

1. Large Behavior Model: A Promptable Digital Twin of the Retail CustomerPrimary source - arXiv - research preprint - Jul 10, 2026

   Used for: The July 10 revised preprint posting, model description, dataset scope, evaluations, and limitations.

   Open source

2. FTC Surveillance Pricing 6(b) Study: Research Summaries; A Staff PerspectiveFederal Trade Commission - government staff report - Date not listed

   Used for: Separate context on granular consumer data, individualized discounts, and pricing.

   Open source

3. AI RMF PlaybookNational Institute of Standards and Technology - government guidance - Date not listed

   Used for: Privacy-risk context for AI-enabled inference.

   Open source

4. Bayesian Predictive Profiles with Applications to Retail Transaction DataAdvances in Neural Information Processing Systems - conference paper - Date not listed

   Used for: Background that historical retail transactions can support individualized predictive profiles.

   Open source