RecHarness pairs bandit routing with LLMs to automate recommender model tuning
The arXiv paper splits automated recommender optimization in two: a bandit router picks the modification direction, while an LLM formulates the concrete code change.
In plain words
Recommender systems are usually improved by hand, with engineers repeatedly adjusting model designs and training setups. RecHarness automates that trial-and-error work: a statistical router chooses which kind of change to try next, and a large language model writes the specific hypothesis and code edit. The authors report steadier gains than letting the LLM steer the whole search.
What was announced
The authors argue that when a large language model both selects modification directions and formulates concrete hypotheses, the search becomes unstable under limited trial budgets. RecHarness therefore decouples the process: a bandit router chooses the next modification direction based on historical validation feedback, and the LLM then generates a concrete optimization hypothesis plus an executable code edit within that direction. A jump-basin mechanism activates a structural-jump arm when local edits stagnate, which the paper says sustains long-horizon exploration.
Reported results come from the paper itself. Across multiple recommendation tasks, datasets, and model backbones, the authors say RecHarness achieved more consistent improvements and used limited trial budgets more effectively than LLM-directed search. In a 7-day online A/B test on a large-scale short-video advertising platform, the selected candidate improved ADVV by 2.084%, Revenue by 0.534%, and Exposure by 0.559%. The code is available on GitHub. These figures reflect the authors' own tests and have not been independently verified.
What it does not show
- Evidence is limited to the paper's abstract, so experimental setup, baselines, and statistical details cannot be checked from the supplied passages.
- The A/B test metrics are the authors' own reported results; no independent confirmation is available.
For security and IT teams
- Status
- The paper reports code availability at the project's GitHub repository, per the abstract.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.