RAS method boosts Cypher query accuracy by 41-50% using error feedback

A new paper on arXiv presents RAS (Reflection-Augmented Scaling), a technique that leverages inference-time compute more efficiently for structured query generation. The researchers focus on Text2Cypher, where language models produce Cypher queries for Neo4j property graph databases. A key insight: non-executable queries (syntax errors) yield system error messages that are typically discarded, but RAS retains and uses them as in-context learning signals for subsequent attempts. This contrasts with Independent Scaling (IS), which simply resamples without memory. Across five code-specialized models and three datasets, RAS cuts the Query Execution Error Rate by 41–50% at n=5 attempts, compared to 32–38% for IS—a 9–12 percentage point improvement.

Beyond raw error reduction, the work highlights a deeper principle: execution errors are not just failures but actionable feedback that can guide inference-time compute allocation. By structuring resampling around prior execution traces, RAS achieves higher executable query rates without additional training or larger models. The method is particularly relevant for enterprise applications relying on natural language interfaces to graph databases, where query correctness directly impacts downstream analytics. The authors suggest RAS could extend to other structured languages (SQL, SPARQL) and other failure modes, paving the way for more robust AI-assisted data interaction.