AgenticRecTune: Multi-Agent with Self-Evolving Skillhub for Recommendation System Optimization
Five specialized agents use Gemini to auto-tune recommendation pipeline configurations.
Modern large-scale recommendation systems are built as multi-stage pipelines (pre-ranking, ranking, re-ranking). While traditional research focuses on optimizing individual models (e.g., improving pre-ranking model structure or training algorithms), system-level configuration optimization is crucial but highly challenging. Any model modification requires new optimal configurations, each stage has different contexts and targets, and success depends on balancing competing online metrics while aligning with shifting production goals. Manual tuning is time-consuming and domain-expertise intensive.
AgenticRecTune, submitted to arXiv by Xidong Wu and nine other authors, introduces an agentic framework of five specialized agents powered by Gemini LLMs. The Actor Agent proposes multiple candidate configurations, the Critic Agent filters suboptimal ones, the Insight and Skill Agents collaborate on a self-evolving Skillhub that summarizes historical results to extract underlying mechanics of each task, and the Online Agent autonomously prepares and runs A/B tests, capturing results. This end-to-end automation dramatically reduces tuning effort and enables continuous optimization as models evolve.
- Five specialized agents (Actor, Critic, Insight, Skill, Online) handle different aspects of configuration optimization.
- Leverages Gemini LLMs to explore optimal configuration spaces for multi-stage recommendation pipelines.
- Self-evolving Skillhub extracts mechanics from historical results and updates skills automatically.
Why It Matters
Automates the tedious manual tuning of recommendation system configurations, enabling faster iteration and better performance.