Human-level Text-to-SQL achieved with verified data and RL

Researchers have developed a new method for Text-to-SQL that achieves human-level accuracy by fine-tuning LLMs with reinforcement learning on verified data, bypassing complex pipeline engineering. They created BIRD-Platinum, a dataset with corrected annotations, which improved Qwen3-235B performance significantly. Further enhancements were made with ReViSQL-BIRD, a reward shaping technique that combines result-based rewards with SQL equivalence verification and external knowledge utilization, enabling Kimi-K2.6 to reach 92.96% accuracy on the Arcwise-Plat benchmark. AI

IMPACT This research demonstrates a path to human-level performance in Text-to-SQL without complex pipelines, potentially simplifying database interaction for users.

RANK_REASON Academic paper detailing a new method and benchmark results for Text-to-SQL. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Arcwise-Plat
• BIRD-Platinum
• BIRD Train
• Kimi K2.6
• Qwen3 235B
• ReViSQL-BIRD
• SPIDER2: A Package to Predict Secondary Structure, Accessible Surface Area, and Main-Chain Torsional Angles by Deep Neural Networks.
• Yuxuan Zhu

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →