Alibaba's RecGPT-V3 boosts Taobao revenue by 3.97% while slashing compute costs
RecGPT-V3 cuts token costs by 200x and resource consumption by 52.4% in production at Taobao.
Large language models are transforming recommender systems, but scaling them introduces inefficiencies. Alibaba's Taobao addresses three key challenges in its latest RecGPT-V3: stateless behavior modeling, a tag-to-item information bottleneck, and inefficient explicit reasoning. The system uses a Memory Hub that distills long-horizon user behavior into condensed units, cutting user-modeling computation by 55.8%. A Hybrid-modal Foundation Model jointly reasons over text tags and Semantic IDs (SIDs), opening a high-bandwidth channel into the item space. Latent Intent Reasoning internalizes verbose chain-of-thought rationales into compact learnable tokens, reducing output token cost by 200x while still producing readable explanations.
Deployed in Taobao's "Guess What You Like" feed, RecGPT-V3 delivers significant online gains: IPV (item page views) up 1.28%, click-through rate up 1.00%, transaction conversion up 1.97%, and gross merchandise value (GMV) up 3.97%. At the same time, end-to-end serving resource consumption drops by 52.4%, making LLM-based recommendations commercially viable at scale. This combination of improved user experience, revenue growth, and dramatic cost reduction sets a new benchmark for production recommender systems.
- Memory Hub reduces user-modeling computation by 55.8% by distilling long-horizon behavior into condensed units.
- Latent Intent Reasoning internalizes chain-of-thought rationales into compact tokens, lowering output token cost by 200x.
- Deployed on Taobao's feed, RecGPT-V3 boosts GMV by 3.97% while cutting end-to-end serving resource consumption by 52.4%.
Why It Matters
Alibaba proves LLM-based recommenders can be both powerful and cost-efficient, driving revenue gains with drastically lower compute.