# Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL

The fine-tuned model scores 92.97% on Arcwise-Plat-SQL with 16-sample voting at $0.56 per task, exceeding the 92.96% human benchmark without multi-stage agent scaffolding.

- Fine-tuned Kimi-K2.6 using reinforcement learning with verifiable rewards on BIRD-Platinum, a newly cleaned dataset that corrects labeling errors across benchmark queries.
- Uses VeriEQL bounded verification to reward true semantic equivalence rather than single-instance execution matches.
- Applies rule-based process rewards requiring the model to generate explicit constraint and verification blocks for provided external knowledge.
- Greedy single-sample decoding achieves 91.37% accuracy at $0.035 per task, an 8.4-point gain over OpenSearch at 37% lower cost.
- Code, cleaned datasets, and training recipes are open-sourced on GitHub at github.com/uiuc-kang-lab/ReViSQL.

## Why it matters

Engineers building natural-language database query tools can achieve human-level SQL generation without complex multi-call orchestration frameworks.

## Sources

- [Thinking Machines Lab: Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL](https://thinkingmachines.ai/news/putting-task-expertise-into-rl)

---

Summarized by dstilled on 2026-08-27. https://dstilled.ai/story/3b417755-a924-4842-9baa-23ffed6559c1