Data-aware candidate selection in NL2SQL translation via small separating instances

Researchers proposed a data-aware candidate selection method for NL2SQL translation that uses separating instances and provenance.

They prototyped the approach and evaluated it on BIRD-DEV against three baselines.

The method performed especially well when there were only two or three candidate SQL queries.

Results also showed strong gains when no consistency score was available, a common low-information setting.