RecHarness is an agentic framework for automating recommender-model optimization. It separates search-direction selection from hypothesis and code generation: a bandit router chooses among modification directions using historical validation feedback, while an LLM proposes a concrete optimization and executable code edit within the selected direction. A structural-jump arm is activated when local edits stagnate, supporting longer-horizon exploration. The paper reports more stable gains and better use of limited trial budgets than LLM-reasoning search across recommendation tasks, datasets, and backbones, and mentions a seven-day online A/B test on a large-scale short-video advertising platform. The supplied abstract is truncated, so the online-test metrics are unavailable here.
No heat snapshots are available in the last 24 hours.