Researchers have developed MARS (Multi-stage Agentic Recommendation System), a novel approach that extends personalization beyond individual prediction stages to orchestrate entire recommendation pipelines. A fine-tuned Qwen3-0.6B model acts as an agent, generating per-user configurations for pipeline mode, retrieval quotas, model selection, and fusion weights. This system was trained in three stages, including Format-Anchored Preference Optimization (FAPO) and Group Relative Policy Optimization (GRPO), and was tested on Taobao with over ten million daily users. A/B tests showed significant gains in key metrics like V2P and GMV, with some scenarios reaching over 7% GMV increase, while also demonstrating reduced serving latency. AI
IMPACT This system could enable more dynamic and personalized user experiences in large-scale e-commerce platforms.
RANK_REASON Paper introducing a new system (MARS) and its training methodology. [lever_c_demoted from research: ic=1 ai=1.0]
Read on arXiv cs.IR (Information Retrieval) →
- ActionMask
- alphaXiv
- CatalyzeX Code Finder for Papers
- Connected Papers
- DagsHub
- Fapou Semebene
- Gotit.pub
- Grpo
- Hugging Face
- Litmaps
- MARS
- Qwen3-0.6B
- ScienceCast
- scite Smart Citations
- Taobao
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →