Taobao uses a small Qwen3 model to set each shopper's recommendation pipeline
Alibaba's Taobao team reports that a fine-tuned Qwen3-0.6B, choosing pipeline settings per user ahead of time, lifted purchases and clicks on five live shopping surfaces. The report is a preprint and has not been peer reviewed.

AI illustrationWhat the study shows
Large shopping sites recommend products in stages. A first stage, called retrieval, pulls thousands of candidates from several sources. Later stages score, trim and reorder them into the short list a shopper sees. Engineers normally fix how these stages are set up, and the same setup applies to every shopper on a given page.
The Taobao team built MARS, a system that gives each shopper their own setup. A small language model, Qwen3-0.6B, reads a summary of the shopper's behaviour and the page they are on. It then writes a structured configuration with seven parts. These include how many stages to run, which candidate sources to use and how many items each supplies, which scoring models to use, and how to weight clicks, purchases and price.
The model runs near-line, meaning in the background ahead of time. Its output is checked, then stored in a cache. The live system reads the latest valid entry and never waits for the model. If no valid entry exists, it falls back to the normal setup. A master switch can turn MARS off for all traffic at once. Between 85% and 95% of live requests used a personalised setup, depending on the page.
The model was trained in three steps. First, a much larger Qwen model wrote example answers, which taught the small model the output format. Valid outputs rose from 1% to 99.8%. Second, a preference step taught it which settings did better, using the authors' own method, FAPO. Third, a reinforcement learning step, where a model learns from rewards, let it try new settings in a replay sandbox scored by a separate prediction model, without showing them to real shoppers.
The authors report two-week A/B tests, where 5% of traffic got MARS and another 5% got the existing setup. On three purchase-focused pages, the share of visitors who bought rose by 2.2% to 5.2%. Sales value rose by 4.8% to 7.3%. On two click-focused pages, the share of users who clicked rose by about 1.1% and 2.3%.
About a fifth of requests were sent through a shorter two-stage pipeline, mostly for visitors who seemed to be passing through. The authors say the deployment was linked to a 12 millisecond lower average response time on the live path. They report the system has run for three months, serving more than ten million daily users.
Our reading
In our reading, the interesting move is where the AI sits. Earlier work used language models to score or retrieve items. Here the model acts more like an operator, choosing which existing tools to run and how to tune them, from a tightly bounded menu. Every setting must stay within a small range around the engineers' defaults. That design makes the outputs checkable before they reach anyone.
The serving design is the part most teams could copy. Generating settings in the background, validating them twice and falling back to a known default keeps a slow model off the path that shoppers feel. A concrete example: a shop could have the model decide overnight that a returning bargain hunter gets wider candidate sourcing and more weight on price, while a visitor who only clicks through a promotion gets a cheaper short pipeline.
On the scale we use, we rate this II, Benchmarked, with care. The evidence is controlled live tests against the production setup, not a public benchmark. The operator reports production use itself, which would support a higher level, but nothing is released, so others cannot check or reuse it.
Read the rest of this story
Membership is free. You get every article in full and the Issue every Sunday.
We send a sign-in link, no password needed. Already a member? Sign inReadiness assessment
| Level | Status | Evidence |
|---|---|---|
| I Lab only | Met | The method is fully described: architecture, three-stage training, reward design, action space, and hyperparameters. A synthetic prompt and action example are included. |
| II Benchmarked | Met | Reports controlled online A/B tests (5% vs 5% traffic, two weeks, intention-to-treat) against the production baseline on five Taobao scenarios, with nominal 95% intervals. It also compares against a DNN baseline and against ORPO. These are internal production tests, not a public benchmark, so this is a loose reading of 'benchmarked'. |
| III Open source | Not yet | No code, model weights, data or repository link is given. Prompts are only excerpted and model and scene identifiers are placeholders. |
| IV In production | Met | Taobao (Alibaba) reports three months of continuous operation serving over ten million daily active users, with cache coverage of 85% to 95%. This is a self-report by the operator in a non-peer-reviewed report, with no independent verification. The report also says the evaluated deployment is associated with a 12 ms lower mean serving-path latency. |
What would move it up
Level 3 needs released code, trained weights or a reproducible environment, even a stripped-down version with public data. For stronger evidence at level 2 and 4, the team should publish stage-by-stage ablations with repeated runs and intervals, break out Minimal-mode routing effects from the personalisation effect, and report the full cost including near-line LLM inference and retraining.
The study
Tong Liu, Chaoqun Hou, Longxiang Xiong et al.. MARS Technical Report. arXiv, 18 November 2026. Read the paper
Notes
- Paper: MARS Technical Report, MARS Team (Taobao, Alibaba), arXiv preprint, 18 November 2026. Not peer reviewed. Source
- No code, weights or data are released with the report.
- Main A/B results (Table 5, relative to production, nominal 95% intervals): Baiyibutie V2P +5.22% [+4.58, +5.86], GMV +7.34% [+6.18, +8.51]; Miaosha V2P +3.88% [+2.62, +5.15], GMV +4.83% [+4.15, +5.50]; ShangouPostPay V2P +2.24% [+0.10, +4.43], GMV +7.17% [+1.56, +12.77]. Table 6: MarketingBlock uCTR +1.08%, pCTR +0.99% [+0.08, +1.90]; Dacu uCTR +2.27%, pCTR +3.47%.
- V2P is the share of exposed users who buy; GMV is total transaction value; uCTR is the share of exposed users who click; pCTR is clicks per impression.
- Ablations (single runs, point estimates): training stages on Baiyibutie V2P +3.12% (SFT), +4.80% (FAPO), +5.22% (GRPO); Miaosha V2P +1.77%, +1.66%, +3.88%. Without expert prompt knowledge, Miaosha V2P −27.00%, GMV −21.23%; with it, +0.20% and +2.40%. MARS vs DNN baseline: Baiyibutie V2P +3.36%, GMV +5.90%; Miaosha V2P +2.03%, GMV +1.76%. FAPO vs ORPO: Baiyibutie V2P +3.70%, GMV +0.32%; Miaosha V2P +1.38%, GMV +3.17%.
- Mode mix: about 79% Full, 1% Streamline, 20% Minimal. Replay agreement across 10,000 requests: mean 92.5%, median 100%. Reward model AUC: 0.76 (clicks), 0.81 (conversions). Retraining weekly on 32 H20 GPUs.
Edited and signed by
Stefan PerdalEditor and publisher86 stories editedStefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.
How this was made
An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.
How we workNo corrections

