Taobao's RecGPT-V3 Recommendation System Tech Report
Bowen Zheng · hf · 2026-07-20
Large language models are shifting recommendation systems from "matching historical behavior patterns" to "reasoning behind intent." To address computational waste, information bottlenecks, and inference latency challenges faced by previous versions at scale, the Alibaba Taobao team launched RecGPT-V3.
Architecture Upgrades & Business Gains:
- Memory Hub: Maintains structured, continuously evolving user memories, distilling long-term behaviors into condensed units, slashing user modeling compute by 55.8%.
- Hybrid Modality Reasoning: Combines natural language and semantic IDs (SIDs) for joint reasoning, bridging open-world knowledge with specific items.
- Latent Intent Reasoning: Internalizes lengthy chains of thought into compact latent tokens, cutting output token costs by 200x while remaining decodable into human-readable explanations.
- Online Performance: Deployed in Taobao's "Guess What You Like" feed, achieving CTR +1.00% and GMV +3.97% in A/B testing, alongside a 52.4% reduction in end-to-end serving resource consumption.
Related event: Alibaba Taobao Unveils RecGPT-V3 Recommendation LLM(2 posts)→
More from Apps
- NVIDIA-backed SimBioSys wins FDA clearance for AI tool predicting breast cancer recurrence at 5 and 10 years — nvidia · 2026-10-27
- Tomo Community launches: crowdsourced recommendations weighted by preference overlap — Jackyhuang · 2026-10-27
- Stripe exec reveals @link balance stack: Privy wallets, Tempo stablecoins, Stripe onramp — jeff_weinstein · 2026-10-27
- A cozy multiplayer city builder made on Sojo that anyone can co-build — jaivinwylde · 2026-10-27
- A dancer's worldline made visible: interactive time-bending demo built with Opus 5.5 — techartist_ · 2026-10-27
- AI may spark a Flash-era renaissance for browser games, says dev — PrasoonPratham · 2026-10-27