RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems
RecHarness limits the LLM’s role, using a bandit router to choose the search direction before code edits are generated.
The paper says that letting an LLM pick both directions and hypotheses can make recommender optimization unstable under tight experiment budgets. RecHarness separates those jobs, routing modification choices from validation history while the LLM proposes the concrete change and executable edit. It also adds a jump-basin mechanism for moments when local edits stall. In tests across recommender tasks, datasets, and backbones, the authors report steadier gains than LLM-reasoning search, plus a 7-day online A/B lift in ADVV, revenue, and exposure. HF Daily Papers' note
The paper says that letting an LLM pick both directions and hypotheses can make recommender optimization unstable under tight experiment budgets. RecHarness separates those jobs, routing modification choices from validation history while the LLM proposes the concrete change and executable edit. It also adds a jump-basin mechanism for moments when local edits stall. In tests across recommender tasks, datasets, and backbones, the authors report steadier gains than LLM-reasoning search, plus a 7-day online A/B lift in ADVV, revenue, and exposure. HF Daily Papers' note
score 4