LCM-Promo: An LLM-Based Uplift Model for Promotion Allocation
A finetuned commerce LLM learns monotonic promotion-response curves from production logs. Its A/B test raised conversion 1% and cut promotion cost 6%.
The claim
A finetuned commerce LLM allocates promotions from monotonic response curves learned on production logs.
The problem
Promotion budgets must reach the users who will actually react to them.
Uplift differs per user, per session, and per time of day.
Generic models ignore session context such as day and delivery shift.
How it works
LCM-Promo is a task-specific Large Commerce Model finetuned on transactional commerce data.
It predicts user intent together with the response to a promotion.
Single-token outputs act as conversion-threshold probabilities for each user.
Monotonic promotion-response curves are learned directly from production logs.
Session day and delivery shift condition every prediction.
Figures
What changes
Conversion rose by 1% in a churn reactivation A/B test.
Promotion cost fell by 6% against the Segment-EMA production baseline.
The comparison ran in production as an A/B experiment.
Limits
A single production A/B test carries all reported results.
The paper reports no test duration, sample size, confidence interval, or segment breakdown.
The reported results come from commerce promotions rather than from financial trading.
Source · arXiv