MOSAIC Picks a GraphRAG Traversal Policy per Query and Beats the Best Fixed Policy by 9.96 Points
arXiv 2609.11065 (submitted 10 Sep 2026) attacks the structural mismatch in GraphRAG, where direct facts want compact local neighborhoods, comparisons want balanced coverage of several targets, and mediated questions need deeper paths through weak connectors, yet most systems traverse the same way for all of them. Mosaic is training-free: an LLM analyzer converts each query's evidence requirements into a bounded policy over seed selection, traversal, stopping and evidence selection while graph, indexes, scoring and generator stay fixed. On GraphRAG-Bench it hits 76.97 query-weighted Answer Correctness on Medical and 64.33 on Novel, 5.13 and 4.43 points over the best prior results, and beats the strongest fixed narrow/medium/wide policy by 9.96 points.
↳ Follow the thread