Online deployment demonstrates improved diversity and conversion in e-commerce users, suggesting that intent-driven models foster sustainable recommendation ecosystems.
Key Points
Develop and deploy an intent-centric recommender framework powered by large language models to overcome the filter bubbles and narrow preferences caused by standard log-fitting systems.
Integrated large language models into user interest mining, item retrieval, and explanation generation pipelines.
Trained the framework using reasoning-enhanced pre-alignment and self-training evolution guided by a cooperative Human-LLM judge system.
Deployed the model into production on the Taobao App to evaluate live performance across multiple user, merchant, and platform metrics.
Enhanced user experience, increasing content diversity (CICD) by +5.53% and dwell time (DT) by +4.35%.
Boosted merchant exposure and transactions, with click-through rate (CTR) up +5.65%, item page views (IPV) up +7.55%, and daily active consumers (DCAU) up +2.46%.
Sustained platform loyalty by improving 30-day user retention (LT-30) by +1.63%.