RecHarness automates AI recommender optimization with 2% ad lift
New AI harness boosts ad revenue by 0.5% and cuts manual tuning time by 90%
RecHarness is a bandit-routed agentic harness for automated recommender-model optimization. It separates the process into two steps: a bandit router picks the next modification direction from historical validation feedback, and an LLM generates a concrete hypothesis and executable code edit within that direction. Across multiple tasks, datasets, and model backbones, it reports more stable performance improvements and more effective use of limited trial budgets than LLM-reasoning 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%. Code is available via the link in the article.
- RecHarness uses a bandit router + LLM agent to automate recommender optimization, cutting manual engineering time by 90%
- In a 7-day A/B test, it improved ad value by 2.084%, revenue by 0.534%, and exposure by 0.559%
- Features a 'jump-basin' mechanism to escape local optima and sustain long-horizon exploration
Why It Matters
RecHarness slashes the time and cost of tuning AI recommenders while delivering measurable revenue gains for ad platforms.