Brief

Researchers have developed MemRerank, a novel framework designed to improve personalized product reranking in LLM-based shopping agents. This system distills a user's extensive purchase history into concise, query-independent signals, addressing issues of noise and relevance mismatch common with raw history. MemRerank utilizes a memory extractor trained with reinforcement learning, demonstrating significant improvements in accuracy, with experiments showing up to a 10.61 absolute point increase in 1-in-5 selection tasks compared to baseline methods. AI