LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM Agent
LuthorSQL handled nearly half of queries locally and cut API costs by 33.3% without reducing accuracy in a 227-question test.
The researchers paired their local 7B model with Claude Opus 5.5, escalating questions when LuthorSQL was uncertain. The test used databases published after Opus’s training cutoff. LuthorSQL generates query-plan scores in one forward pass, executes up to 32 candidate plans and selects an answer with a trained validator. On BIRD’s hidden test set, the standalone model reached 68.03% execution accuracy at 0.84 seconds per question on a single GPU.
HF Daily Papers' note
The researchers paired their local 7B model with Claude Opus 5.5, escalating questions when LuthorSQL was uncertain. The test used databases published after Opus’s training cutoff. LuthorSQL generates query-plan scores in one forward pass, executes up to 32 candidate plans and selects an answer with a trained validator. On BIRD’s hidden test set, the standalone model reached 68.03% execution accuracy at 0.84 seconds per question on a single GPU.
HF Daily Papers' note
score 5