TraceSQL: Traceable Answerability Estimation for Reference-Free Text-to-SQL Verification

Excerpt

arXiv:2608.17795v1 Announce Type: new Abstract: Text-to-SQL systems are commonly evaluated using ground-truth SQL queries or reference execution results, but such supervision is unavailable at inference time in real-world deployments. This creates a critical verification problem: given only a user question, database context, and generated SQL, can a system estimate whether the generated query is likely to correctly answer the question? Recent approaches use LLMs as judge or specialized agents to