Reproduce “TRUST-SQL: Tool-Integrated Multi-Turn Reinforcement Learning for Text-to-SQL over Unknown Schemas” on OpenResearch

About this paper

Ai Jian, Xiaoyun Zhang, Eryu Guo, Wanrou Du, Jingqing Ruan, Jiangbo Pei, Weipeng Zhang, Ke Zeng and 1 more

TRUST-SQL recasts Text-to-SQL as an interactive unknown-schema task: the model must inspect database metadata, identify relevant structure, and then generate and verify SQL rather than receiving the full schema upfront. Its phase-aware reinforcement-learning design enables Qwen3 models to reach up to 68.1% greedy and 70.1% majority-vote execution accuracy on BIRD-Dev, while reducing schema hallucinations to about 2.8% in a pilot study.

Key insights

• The explicit Propose checkpoint sharply reduces schema hallucinations, lowering their share of pilot-study failures from 26.4% to approximately 2.8%, although semantic reasoning remains the dominant challenge.
• Training with selective exploration allows TRUST-SQL to operate without full-schema prefilling: TRUST-SQL-4B reaches 64.9% on BIRD-Dev, compared with 29.3% for the untrained Qwen3-4B model without prefilling.
• Execution-coupled sparse schema rewards outperform uncoupled and dense alternatives, while SFT warm-up prevents the agent from exploiting rewards by querying nearly the entire database.

Summary by alphaXiv.