SIRIUS-SQL: Anchoring Multi-Candidate Text-to-SQL in Execution Feedback

SIRIUS-SQL introduces a multi-candidate approach to Text-to-SQL generation that addresses redundancy, execution error classification, and selector limitations through difficulty-smoothing reinforcement learning, targeted repair mechanisms, and hybrid confidence-gated selection. The system achieves 75.88% accuracy on BIRD dev and 91.20% on SPIDER test, surpassing previous state-of-the-art multi-candidate systems.

SIRIUS-SQL represents a meaningful advance in natural language-to-database query translation, a foundational capability for autonomous data systems and AI agents that need to interact with structured databases. The research tackles a practical problem in production systems: single-pass text-to-SQL generation often fails on complex schemas, making multi-candidate approaches necessary. However, naive multi-candidate systems suffer from diminishing returns and lack nuanced error handling.

The three-pronged solution demonstrates sophisticated system design. The difficulty-smoothing RL approach generates diverse candidates rather than redundant variations, directly addressing the efficiency problem in sampling. The execution-grounded lifecycle that differentiates between runtime errors, timeouts, and empty results reflects understanding that different failure modes require different corrections. The confidence-gated hybrid selector combining execution agreement with structural checks avoids over-reliance on any single voting mechanism.

This work matters because Text-to-SQL reliability directly impacts enterprise AI adoption. Database query generation powers semantic search, automated reporting, conversational analytics, and autonomous data exploration tools—all high-value applications. The benchmark results on BIRD dev and SPIDER test indicate progress on genuinely difficult problems, not incremental gains on easier datasets.

The research validates that ensemble approaches with intelligent filtering outperform single-model solutions. Future work likely involves scaling these techniques to even larger schemas, handling more complex SQL dialects, and reducing latency for real-time applications. Organizations building data-centric AI systems will monitor adoption of these techniques.

• →SIRIUS-SQL achieves 75.88% accuracy on BIRD dev through difficulty-smoothing RL that generates diverse executable SQL candidates.
• →The system classifies execution outcomes (errors, timeouts, empty results) to apply targeted repairs rather than generic corrections.
• →Hybrid selector combining execution-result agreement with structural checks outperforms single-angle voting mechanisms.
• →Two generalist LLM pairings surpass Agentar-Scale-SQL, the previous strongest multi-candidate system on BIRD dev.
• →Research advances Text-to-SQL reliability for enterprise applications including semantic search, analytics, and autonomous data exploration.