Hacker News · AI·· 6 h agoAI score58
Thinking Machines details ReViSQL-K2.6, an RL-trained text-to-SQL model reaching human-level accuracy without scaffolding
Putting Task Expertise into RL Achieves SOTA Performance on Text-to-SQL
AI summary
Thinking Machines describes ReViSQL-K2.6, a Kimi-K2.6 model fine-tuned with RLVR on a cleaned training set, which reaches 92.97% on Arcwise-Plat-SQL with 16-sample self-consistency at $0.56 per task, above the 92.96% human proxy. The team audited 2.5k BIRD Train instances and found errors in 61.1% of them, including incorrect gold SQL in 52.1%, then released the corrected set as BIRD-Platinum. Reward shaping adds VeriEQL equivalence checks and rule-based process rewards for external knowledge, and the model beats GPT-5.6 Sol Ultra and Claude Fable 5 at 12–15% of their cost.
Source:Hacker News · AI · thinkingmachines.ai