Updated ETGet the DI Daily by Email

Technology news for e-commerce professionals, with our take on why it matters.

Agent that tunes Taobao recommendations per user lifted GMV up to 7.34%, paper reports

The MARS system runs a fine-tuned Qwen3-0.6B model outside the serving path, and deployment cut mean latency by 12 ms.

Source: arXiv cs.IR (Information Retrieval)

A technical report posted to arXiv describes MARS (Multi-stage Agentic Recommendation System), which uses an AI agent to choose how a recommendation pipeline runs for each user. The authors say pipeline-level decisions such as channel activation, quotas, model selection and score fusion are usually tuned by engineers and applied the same way to every user within a scenario.

In MARS, a near-line agent built on a fine-tuned Qwen3-0.6B model generates a per-user configuration covering pipeline mode, retrieval quotas, model selection, fusion weights and reranker settings. Configurations are validated, cached and consumed asynchronously, so the cascade never waits for the language model, according to the report. Schema checks, fallbacks and a kill switch guard serving.

The authors report two-week A/B tests across five Taobao scenarios with over ten million daily users, with positive gains on all key metrics. The largest reached +5.22% V2P and +7.34% GMV, with positive click-through lifts on both click surfaces. About 20% of requests use a lightweight Minimal mode, and the deployment showed 12 ms lower mean serving latency than production, the report says.

Read the full article at arXiv cs.IR (Information Retrieval)

Our take

Most commerce teams personalize model scores but set pipeline rules once for everyone, and this report suggests per-user pipeline settings can lift GMV without slowing pages.

  • Results are self-reported from one platform and have not been independently replicated.
  • Ask your recommendation vendor whether quotas and fusion weights can vary by user.
  • Note the asynchronous, cached design: it keeps LLM costs out of the request path.

Was our take useful?

Key takeaways

  1. Authors report up to +7.34% GMV and +5.22% V2P in two-week Taobao A/B tests
  2. Tests covered five Taobao scenarios with over ten million daily users
  3. The agent runs on a fine-tuned Qwen3-0.6B model and cut mean latency by 12 ms