Taobao's RecGPT-V3 cuts LLM recsys tokens 200x, lifts GMV 3.97%

Alibaba's Taobao just published receipts for something a lot of teams have been trying and mostly not talking about publicly, which is running large language models inside a live web-scale product feed without lighting the infrastructure budget on fire. The technical report, posted to arXiv on July 17, 2026, describes RecGPT-V3, the third iteration of a recommendation stack the authors say has already had V1 and V2 deployed in production at Taobao.

The headline numbers are the ones to pay attention to. The paper claims a 200x reduction in output token cost, coming from what the authors call latent intent reasoning, alongside a 52.4% cut in end-to-end serving resource consumption. On the user-modeling side, a Memory Hub that condenses long-horizon behavior into structured units is credited with cutting user-modeling computation by 55.8%. The system pairs natural-language reasoning for open-world knowledge with Semantic IDs for concrete item grounding, so the model reasons in words while pointing at real SKUs.

The result that will get engineering leaders' attention is the live A/B on Taobao's 'Guess What You Like' feed: item page views up 1.28%, click-through up 1.00%, transaction count up 1.97%, and gross merchandise value up 3.97%. On a surface that size, a GMV move of that shape is real money, and it lands alongside the infra savings rather than being paid for by them.

The honest caveat is that this is a self-published technical report from the team that built the system. The paper does not disclose what baseline the A/B was measured against, what absolute cost per query looks like, how the system performs on cold-start users or long-tail items, or the model size and hardware footprint behind the percentages. Take the specifics as reported, not settled.

What is interesting even with those caveats is the direction. For anyone running a large recommendation surface, the pitch has moved from 'can LLMs eventually be economic in the ranker' to 'here is a public architecture and a live number to argue against.' Amazon, Meta, ByteDance and every retailer with an in-house ranking team now have a named reference point to benchmark their own experiments against, which is exactly what has been missing from this conversation until now.