MAC-KGC: A Multi-Agent Collaborative Link Prediction Framework for Knowledge Graph Completion

Abstract

Real-world knowledge graphs inevitably contain missing facts, and inferring these absent links remains a central challenge. While large language models (LLMs) bring strong semantic understanding to this task, current single-model pipelines conflate retrieval, reasoning, and verification into one monolithic process, producing hallucinated entities and shallow inference. We introduce MAC-KGC, a collaborative multi-agent architecture in which three specialized LLM agents work in concert through structured communication and iterative feedback. Evaluations on FB15k-237 and WN18RR show that MAC-KGC surpasses both embedding baselines and single-LLM alternatives across all ranking metrics.