RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems
TLDR
RecHarness uses a bandit router and LLM to automate recommender model optimization, achieving stable improvements and real-world gains.
Reasoning
The paper presents a novel approach combining bandit algorithms with LLMs for automated recommender system tuning, supported by both offline experiments and an online A/B test. Strengths include clear separation of direction selection and hypothesis generation, and a jump-basin mechanism for long-horizon exploration. Weaknesses: limited to recommender systems, not general scientific discovery; no comparison to other automated research agents beyond LLM search.
Read-first score
Read-first score 51.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.
Field roles
Rank sensitivity
Stability: volatile; rank range: 79.
Keyword Scores
Deep Analysis
Innovations
- Bandit-routed optimization that decouples direction selection from hypothesis generation
- Jump-basin mechanism that activates a structural-jump arm to escape local stagnation
Methodology
RecHarness separates automated model optimization into two steps: a bandit router selects the next modification direction based on historical validation feedback, and an LLM generates a concrete optimization hypothesis and executable code edit within that direction. A jump-basin mechanism triggers a structural-jump arm when local edits stagnate to sustain long-horizon exploration.
Key Results
RecHarness achieves more stable performance improvements and uses limited trial budgets more effectively than LLM-reasoning search across multiple tasks, datasets, and model backbones. In a 7-day online A/B test on a large-scale short-video advertising platform, the selected candidate improves ADVV by 2.084%, Revenue by 0.534%, and Exposure by 0.559%.