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.