DPC: Training-Free Text-to-SQL Candidate Selection via Dual-Paradigm Consistency

2026-04-16

AI Summary

This work addresses the challenge that large language models (LLMs) struggle to reliably evaluate the correctness of generated SQL queries in Text-to-SQL tasks, resulting in a gap between high logical accuracy and low execution accuracy. Existing training-free selection methods are often undermined by systematic biases and symbolic blind spots. To overcome these limitations, the authors propose the DPC framework, which reframes SQL selection as a deterministic verification task grounded in observable data. DPC employs a multi-agent architecture: SLICER and TESTER agents collaboratively construct a Minimal Discriminative Database (MDD), while a SOLVER agent cross-verifies the consistency between SQL and Python/Pandas execution results. This training-free approach integrates multi-agent collaboration, MDD construction, and cross-paradigm execution consistency, significantly outperforming existing methods on the BIRD and Spider benchmarks—achieving up to a 2.2% absolute gain in execution accuracy across multiple LLMs and consistently surpassing strong baselines such as Self-Consistency.

Technology Category

• Natural Language Processing: Question Answering
• Search and Optimization: Learning to Search
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

While Large Language Models (LLMs) demonstrate impressive proficiency in generating SQL queries, they fundamentally lack the capability to self-evaluate correctness without an execution oracle. This limitation creates a stark Generation-Selection Gap, where high potential accuracy (Pass@K) fails to translate into execution accuracy (Pass@1). Although supervised verifiers offer mitigation, they incur prohibitive annotation costs and suffer from domain fragility. Consequently, recent research has pivoted to the training-free setting. However, existing methods--such as Self-Consistency or LLM-as-a-Judge--remain hampered by systematic bias (consensus on hallucinations) and symbolic blindness (inability to simulate execution states). We introduce DPC (Dual-Paradigm Consistency), a multi-agent framework that reformulates SQL selection from a probabilistic guessing task on hidden data into a deterministic verification task on visible data. Specifically, DPC employs a SLICER and a TESTER agent to collaboratively construct a Minimal Distinguishing Database (MDD)--an adversarial, fully observable micro-environment engineered to expose logical discrepancies between candidates. To break the self-correction bias, a SOLVER agent then verifies the SQL candidates by cross-referencing their execution against a parallel Python/Pandas solution. By validating execution consistency between declarative (SQL) and imperative (Python) paradigms, DPC robustly discriminates correct logic from systematic hallucinations. Experiments on BIRD and Spider across multiple LLMs demonstrate that our method consistently outperforms existing selection baselines, achieving absolute accuracy improvements of up to 2.2% over strong competitors like Self-Consistency.

Problem

• Text-to-SQL
• Candidate Selection
• Generation-Selection Gap
• Hallucination
• Execution Accuracy

Innovation

• Training-Free
• Text-to-SQL
• Dual-Paradigm Consistency
• Minimal Distinguishing Database
• Execution Verification