CoRe: A Continuously Reward-Finetuned LLM Query Rewriter for Multi-Stage Context-Aware Relevance in Web-Scale Video Search

The article introduces CoRe (Context Relevance), a continuously reward-finetuned LLM query rewriter designed for a short-video search engine, which has been redeployed weekly for over five months. It employs a semi-online Mixed Preference Optimization loop that utilizes a DPO-style pairwise objective for efficient training at scale, leveraging a deployed multimodal relevance model to align training rewards with production ranker consumption. This system significantly reduces change-query rates on impacted queries while improving overall relevance and engagement metrics, making it a valuable tool for practitioners aiming to enhance query rewriting in dynamic data environments.