LuthorSQL: a 7B one-pass text-to-SQL gate that cuts agent API costs 33.3%

Ibrahim Memon · hf · 2026-11-17

UC Irvine researchers present LuthorSQL, a non-autoregressive 7B text-to-SQL model acting as a fast System 1 gate for a slower LLM agent, escalating only uncertain queries. It hits 68.03% execution accuracy on BIRD's hidden test at 0.84s/question on one GPU, transfers zero-shot across five SQL dialects, and in a confidence cascade with Claude Opus 5.5 answers nearly half of queries locally on post-cutoff databases, cutting API cost 33.3% at equal accuracy. Contamination probing shows Opus reproduces 48–92% of public benchmark gold queries verbatim without schemas.

Original post →

More from coding & agent

coding & agent channel →