RecGPT-V3 cuts Taobao serving compute 52.4% while lifting GMV 3.97%
pmttyji · reddit · 2026-07-26
RecGPT-V3 is presented as a stateful, hybrid-modal recommender for Taobao that addresses three scaling problems in earlier RecGPT systems: repeated full-history reprocessing, a lossy tag-to-item bottleneck, and expensive explicit chain-of-thought reasoning.
The system introduces a Memory Hub to maintain structured long-term user memory, cutting user-modeling compute by 55.8%; a hybrid foundation model that reasons over both natural-language tags and Semantic IDs; and latent intent reasoning that compresses verbose rationales into latent tokens while keeping them decodable. Deployed in Taobao’s “Guess What You Like” feed, RecGPT-V3 reportedly lifts IPV by 1.28%, CTR by 1.00%, TC by 1.97%, and GMV by 3.97%, while reducing end-to-end serving resource consumption by 52.4%.
More from Companies & People
- Agency founder says software now means betting 6–24 months ahead while shipping to hundreds of users today — alexgoughcooper · 2026-07-26
- Notion welcomes ZeroEntropy_AI as it pushes a multi-model routing future — CShorten30 · 2026-07-26
- Reservoir Farms says 20+ startups are now paying to test agtech on its farms — DynamicWebPaige · 2026-07-26
- AMD joins NVIDIA–Microsoft open-weight AI letter as industry lines up behind it — firstadopter · 2026-07-26
- DTC CPG founder says Claude replaced GTM vendors that charged thousands — MartinGTobias · 2026-07-26
- San Francisco make-a-thon on July 31 offers $4k top prize for one-night prototypes — floguo · 2026-07-26