RRCM paper boosts LLM recommenders with ranking-driven retrieval and reasoning

Large Language Models (LLMs) are increasingly used for next-generation recommender systems, but current approaches suffer from fixed context construction strategies — they predefine how to incorporate collaborative signals (user behavior) and item metadata, often leading to overloaded or underutilized context windows. The new paper, RRCM (Ranking-Driven Retrieval over Collaborative and Meta Memories), proposes a flexible framework that starts from a lightweight user-history context and learns whether to directly recommend, retrieve collaborative evidence, retrieve item metadata, or interleave both through reasoning. Both memory types are stored as natural language and accessed via a unified retrieval interface, eliminating handcrafted injection or static rules.

To optimize this dynamic retrieval policy, the authors apply group relative policy optimization (GRPO) with a ranking reward that directly reflects final top-k recommendation quality. Experiments show RRCM significantly outperforms traditional baselines and diverse LLM-based recommendation approaches. This work addresses the critical context-efficiency bottleneck in LLM recommenders by letting the model itself decide what evidence matters per instance, rather than relying on heuristic filtering or aggressive compression that can discard fine-grained signals.