SIRIUS-SQL : Anchoring Multi-Candidate Text-to-SQL in Execution Feedback
Leo Luo, Haining Xie, Siqi Shen, Zhipeng Ma, Rui Ling, Hang Xu, Hefeng Jiang, Dingwei Chen, Yang Li, Peng Chen, Jie Jiang · 2026
Read SIRIUS-SQL : Anchoring Multi-Candidate Text-to-SQL in Execution Feedback by Leo Luo, Haining Xie, Siqi Shen, Zhipeng Ma, Rui Ling, Hang Xu, Hefeng Jiang, Dingwei Chen, Yang Li, Peng Chen, Jie Jiang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-01246.txt.
Description
Luo and co-authors anchor text-to-SQL in live execution feedback. One-pass SQL on complex schemas is often wrong. Many systems draft many queries and vote on them. Votes alone miss key flaws. Drafts from one model tend to repeat each other. One fix for all error types fits none well. SIRIUS-SQL attacks all three flaws at once.
A tuned 32B model drafts diverse, runnable SQL. A general model fills gaps the specialist leaves. Each run result is classed by type, then fixed in a targeted way. A hybrid selector weighs result match and SQL form. It scores 75.88% on BIRD dev. It scores 91.20% on the SPIDER test set.
Catalog record
| CREATOR | Leo Luo, Haining Xie, Siqi Shen, Zhipeng Ma, Rui Ling, Hang Xu, Hefeng Jiang, Dingwei Chen, Yang Li, Peng Chen, Jie Jiang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 90 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.01246