Closing the Auto-Research Loop: An AI Co-Scientist for Production Search Ranking
Researchers have developed an AI Co-Scientist framework that integrates LLM agents with direct cloud-compute access to automate the research loop for search ranking systems. This framework utilizes a hybrid agent architecture, employing single-LLM agents for routine tasks and multi-LLM consensus for critical decisions, involving models like GPT-5.2, Gemini Pro 3, and Claude Opus 4.5. The system demonstrated an additional 0.083% gain on top of a transformer baseline, contributing to a total of 0.201% offline improvement in search ranking performance for a travel platform. The AI Co-Scientist also identified and proposed useful techniques from natural language processing and visual perception that were previously absent from the production stack. AI
IMPACT Automates research cycles for AI systems, potentially accelerating development and cross-disciplinary knowledge transfer.