Q-value guided Text-to-SQL generation: Structured reasoning meets efficient inference exploration

doi:10.1016/j.ipm.2025.104607

Q-value guided Text-to-SQL generation: Structured reasoning meets efficient inference exploration

2026-01-15

0

PRE

AI

DOI：10.1016/j.ipm.2025.104607

Original

Original request for help

Save

Abstract

En 中文

• A step-wise chain-of-thought decomposed generation that mimics human step-wise reasoning, significantly improving SQL accuracy. • An MCTS framework enhanced with consensus filtering, which suppresses misleading intermediate outputs during Q-value training. • A memory-efficient Q-value network that enables seamless integration with LLM decoding, achieving fast inference without sacrificing performance.

Q-value guided Text-to-SQL generation: Structured reasoning meets efficient inference exploration

Abstract

Journal

Organization