What's this about?
Alibaba's technical report on RecGPT-V3, the third generation of Taobao's LLM-based recommender system — moving recommendation from matching historical co-occurrence patterns ("people who bought X bought Y") to reasoning about the *intent* behind behavior, deployed at one of the world's largest e-commerce scales (HF 7↑).
What exactly changed vs. a normal recommender?
Classical recommenders are pattern-matchers over clicks and purchases. An LLM-based recommender instead tries to model *why* — inferring "user is furnishing a first apartment" from browsing, then reasoning about what that project needs next. V1 centered user understanding, V2 scaled it with coordinated multi-agent components, V3 is the production-hardened iteration.
Why does it matter?
Production reports from systems serving hundreds of millions of users are rare; they show which LLM-reasoning ideas survive contact with latency budgets and real revenue metrics.