# GRASP preprint reports retrieval gains by combining text and graph searches

_Published Friday, October 16, 2026 at 10:29 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report that GRASP, a system for searching knowledge bases containing text and relationships, improved average Hit@1 by 18.6 points over the strongest baseline on three STaRK benchmarks. The preprint reports a query time of 2.47 seconds.

A small language model turns each query into an executable plan for following relationships. GRASP combines the resulting ranking with a text-search ranking, then uses a fine-tuned model to reorder the candidates. Reinforcement learning trains the planner to improve the combined ranking rather than its own results alone. The reported gains concern retrieval across the three benchmark datasets.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2605.30237)

---
Canonical: https://techandbusiness.org/newswire/nNRstHQXzXknMUt1NHw2Rr
Published: 2026-10-16T14:29:22.809Z
Story chronology: 2026-10-16T04:00:00.000Z
Retrieved: 2026-10-16T17:02:11.967Z
Publisher: Tech & Business (techandbusiness.org)