CG-MAR: LLM-Based Multi-Agent Recommendation System via Community Graph

Abstract

Large language models (LLMs) are propelling recommender systems toward generative paradigms and multi-agent collaboration by using natural language as a unified interface. However, relying solely on prompt engineering is insufficient to capture the temporal dynamics of user preferences. Therefore, it is necessary to explicitly incorporate temporal information into prompts and fuse evidence from multiple sources. Existing approaches lack a traceable, deterministic workflow and do not yet provide an interpretable, unified fusion of community-graph-based scoring, topic-aware TF-IDF content representations, priors over next-hop transition sequences, and session-wise co-occurrence patterns. Moreover, their limited cross-domain reusability often leads to hallucinations and instability. To address these limitations, we propose CG-MAR, a community-reasoning-enhanced multi-agent recommendation framework driven by LLMs. CG-MAR adopts a candidate closed-loop strategy and a deterministic five-agent pipeline consisting of retrieval, listwise reranking, pairwise tournament refinement, verification, and memory. Experimental results demonstrate that CG-MAR consistently improves the performance of recommendations. We further compare CG-MAR against eleven baselines on three evaluation metrics, and the results show that CG-MAR outperforms all competing methods.