Kuaishou's RecHarness: Automating Recommender Iteration with LLMs
_reachsumit · x · 2026-08-03
Kuaishou proposed the RecHarness framework, leveraging Large Language Models (LLMs) to automate the optimization of recommender models, reducing manual trial-and-error costs.
- Decoupled Design: A bandit router selects the optimization direction based on historical feedback, while the LLM generates concrete optimization hypotheses and code edits.
- Escaping Local Optima: Introduces a jump-basin mechanism to activate structural jumps when local edits stagnate, sustaining long-horizon exploration.
- Business Impact: In a 7-day online A/B test, the selected candidate improved ad revenue by 0.534% and exposure by 0.559%.
More from Research
- Dally 2022 Model: SRAM Access Energy Varies by Two Orders of Magnitude — jwt0625 · 2026-08-03
- AI Paper Caught Plagiarizing, Cites 34-Page Claude-Generated Slop — suchenzang · 2026-08-03
- Open-source tool turns coding agents into one-command autonomous research paper generators — tom_doerr · 2026-08-03
- Multi-Agent Emergence: Decentralized Swarm Paints with Coordinated Trails — johnowhitaker · 2026-08-03
- Awesome AI Memory: A Curated Knowledge Base for LLM & Agent Memory — tom_doerr · 2026-08-03
- Google's Science One Framework: Verifiable Autonomous Research Agents — maier_ak · 2026-08-03