RecHarness: Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems
A recent study published on arXiv (2607.29241) presents RecHarness, a Bandit-Routed Agentic Harness aimed at automating the enhancement of recommender models. This system tackles the issue of dependence on engineers for manual adjustments to architecture, objectives, and training strategies. RecHarness divides the optimization into two phases: a bandit router determines the next modification direction based on past validation results, and a large language model (LLM) formulates a specific optimization hypothesis along with an executable code change. To facilitate long-term exploration, RecHarness integrates a jump-basin mechanism that triggers a structural-jump arm when local modifications become ineffective. The findings indicate that RecHarness consistently surpasses existing automated machine learning (AutoML) and LLM-based optimization benchmarks, achieving better results with reduced computational expenses. The research underscores the promise of merging bandit algorithms with LLM-driven code generation for developing self-evolving recommender systems capable of adapting autonomously.
Key facts
- RecHarness is a Bandit-Routed Agentic Harness for automated recommender model optimization.
- It separates optimization into a bandit router and an LLM generating hypotheses and code edits.
- The jump-basin mechanism activates a structural-jump arm when local edits stagnate.
- RecHarness outperforms existing AutoML and LLM-based baselines on multiple recommendation benchmarks.
- It achieves superior performance with lower computational costs.
- The paper is available on arXiv with ID 2607.29241.
- The research addresses the need to reduce manual iteration by engineers in recommender model optimization.
- The system uses historical validation feedback to guide the bandit router.
Entities
—