[arXiv]score: 0.15
GRASP Lifts STaRK Benchmark Hit@1 from 62.0 to 73.9 on Semi-Structured Knowledge Bases
May 29, 2026
GRASP combines plan-guided graph retrieval, plan-conditioned dense fusion, and a fine-tuned reranker to improve retrieval over semi-structured knowledge bases. It advances state-of-the-art Hit@1 by 11.9 points averaged across all three STaRK benchmarks covering product search, academic search, and precision medicine.
cs.IRcs.CLcs.LG
HOW THIS AFFECTS YOU
●
builderIf you're building RAG pipelines over typed entity graphs, GRASP's architecture offers a concrete blueprint with measurable gains over standard hybrid retrieval without requiring end-to-end fine-tuned traversal models.
●
researcherThe three-stage ablation confirms each component contributes independently — plan-based graph retrieval and reranking are the highest-leverage additions over baseline hybrid systems.