KoRe: Compact Knowledge Graph Tokens Achieve 10x Token Reduction for LLM Grounding

arXiv·medium signalXBlueskyLinkedInCopy link

KoRe encodes 1-hop knowledge graph sub-graphs into compact discrete tokens and injects them into an LLM backbone, achieving competitive results across three benchmarks while using up to 10x fewer tokens than standard RAG approaches. This addresses hallucination, knowledge opacity, and update difficulty by making the knowledge source explicit and editable. For practitioners running KG-backed RAG systems, this is a practical path to dramatic cost reduction.