Kuaishou's RecHarness: Automating Recommender Optimization with Bandit-Routed LLM Agents

kuaishou · hf · 2026-08-04

Kuaishou's team proposes RecHarness, a bandit-routed agentic harness designed to automate the optimization iteration of recommender models.

Pain Point: Optimizing traditional recommender models relies heavily on engineers manually tweaking architecture, objectives, and training strategies. Letting LLMs both select directions and generate code often leads to unstable search under limited budgets.
Solution: RecHarness separates optimization into two steps: a bandit router selects the next modification direction based on historical validation feedback, while the LLM generates a concrete optimization hypothesis and code edit within that direction. A jump-basin mechanism is introduced to activate structural jumps when local edits stagnate.
Results: In a 7-day online A/B test on a large-scale short-video advertising platform, the method improved ADVV by 2.084%, Revenue by 0.534%, and Exposure by 0.559%.

More from coding & agent