Reproduce “CypherGuard: Database-Grounded Correction for Reliable Text-to-Cypher” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zebin Chen, Dong Wen, Makbule Gulcin Ozsoy, Wenjie Zhang

CypherGuard is a training-free correction layer for Neo4j Cypher, combining rules mined from training-graph failures, deterministic schema and storage-statistic checks, and execution-triggered repair. On CypherBench, with gpt-5.4-mini, it raises execution accuracy from 63.42% to 88.50% across 2,348 questions on seven graphs. This inference-time result does not show model training or faster computation; plausible but incorrect rows can still escape correction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.