Test-Time Verification for Text-to-SQL via Outcome Reward Models

Short summary

Researchers introduce GradeSQL, a framework using Outcome Reward Models (ORMs) to verify Text-to-SQL outputs at inference time, replacing heuristic selection strategies. On BIRD and Spider benchmarks, ORM-based selection outperforms execution-based approaches by up to 4.33%, with stronger gains on complex queries. Code and datasets are publicly available.

• •GradeSQL uses learned semantic scoring (ORMs) instead of heuristic signals for Text-to-SQL verification
• •Achieves +4.33% on BIRD and +2.10% on Spider benchmarks over baseline approaches
• •Scales effectively with larger candidate sets and includes public code/datasets

Generated with AI, which can make mistakes.