RRCM: Ranking-Driven Retrieval for LLM-Based Recommendation

A new paper on arXiv (2605.07129) introduces RRCM, a ranking-driven retrieval framework that enhances LLM-based recommenders by integrating collaborative behavioral evidence and item-side metadata through collaborative and meta memories. The approach addresses key challenges in constructing decision-relevant contexts from heterogeneous evidence, overcoming fixed context strategies and context-efficiency bottlenecks that plague existing methods. RRCM dynamically retrieves the most beneficial information for each recommendation instance, improving accuracy and efficiency.