Reproduce “MOSAIC: Query-Aware Exploration Policy Adaptation for GraphRAG” on OpenResearch

About this paper

EunKyeong Lee, Kyeong-Jin Oh, Jinwon Kim, Hye Woo Lee, Minsang Song, Hyeongjun Jang, Junyoung Youn

The MOSAIC system adapts graph retrieval to each question by having an LLM choose bounded settings for starting nodes, traversal, stopping, and evidence retention, while keeping the graph and answer generator shared. On the Medical benchmark, it achieves 76.97% Answer Correctness, versus 67.01% for Fixed Medium, but takes 9.68 seconds versus 5.29, so it reduces graph-search effort, not elapsed time.

Key insights

• Query adaptation is most useful when it changes how the graph is explored, not merely which nodes receive higher relevance scores.
• Policy choices align with independently identified question properties: multi-target queries trigger two-view exploration, multi-hop queries receive more connector search, and local questions are more often assigned shallow traversal.
• More retrieval is not inherently better: MOSAIC improves accuracy while evaluating 81.9% fewer paths and retaining 47.2% fewer evidence items than a fixed wide policy.

Summary by alphaXiv.