Identifying Adjustable Units for Self-Improving Recommender Systems

Ascertain which modules within large-scale industrial recommender system pipelines (including recall, ranking, re-ranking, and strategy-control components) can autonomously adjust their internal structure or interactions as part of explicit mechanisms for system-level self-improvement.

Background

The paper argues that conventional multi-stage recommender pipelines remain static and rely heavily on manual tuning, limiting their ability to adapt to heterogeneous data and multi-objective constraints. In discussing the lack of continual, autonomous improvement, the authors explicitly note uncertainty about which parts of the system should be capable of adjusting themselves.

This uncertainty motivates the shift to an agentic formulation, where certain functionally closed and independently evaluable units become agents with evolvable decision spaces. Clarifying which units should be capable of autonomous structural or interaction-level adjustments is presented as an unresolved aspect in enabling explicit self-improvement mechanisms.