Replacing Training with Memory: Listwise Selection for Text-to-SQL
MaP-SQL uses retrieved “memories” at inference time to choose among SQL candidates without fine-tuning a selector.
The paper replaces learned listwise selection behavior with structured examples distilled from training data. Those memories guide how a question maps to schema elements, SQL operations, and expected outputs. It also reduces ordering bias by aggregating rankings across multiple candidate permutations. On BIRD-dev, the method beats R^3-SQL by 2.02 execution-accuracy points using the same candidate sets, while using 2.92x fewer tokens. HF Daily Papers' note
The paper replaces learned listwise selection behavior with structured examples distilled from training data. Those memories guide how a question maps to schema elements, SQL operations, and expected outputs. It also reduces ordering bias by aggregating rankings across multiple candidate permutations. On BIRD-dev, the method beats R^3-SQL by 2.02 execution-accuracy points using the same candidate sets, while using 2.92x fewer tokens. HF Daily Papers' note
score 4