EviLink: Multi-Path Schema Linking with Uncertainty-Guided Evidence Acquisition for Large-Scale Text-to-SQL

2026-05-28 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

EviLink is a novel approach to schema linking in large-scale Text-to-SQL systems, designed to address the challenge of identifying compact yet sufficient schema context from extensive and ambiguous databases. Unlike existing methods that treat schema linking as a deterministic selection along a single SQL path, EviLink reframes it as uncertainty-aware schema-need inference across multiple plausible SQL paths. This system distinguishes required schema items from path-dependent uncertain ones, acquiring evidence only when necessary. EviLink instantiates this by combining multi-hypothesis schema grounding with uncertainty-guided evidence acquisition. Experimental results on BIRD-Dev and Spider2-Snow datasets demonstrate that this perspective improves the balance among schema completeness, schema relevance, and token cost. Specifically, on Spider2-Snow, EviLink achieves a 90.15% field-level strict recall rate and utilizes 123.30K average tokens, leading to improved downstream SQL generation with a fixed generator.

Key takeaway

For NLP Engineers developing Text-to-SQL systems for large, ambiguous databases, consider adopting EviLink's multi-path schema linking approach. Your current deterministic, single-path methods may be suboptimal for complex queries. By reframing schema linking as uncertainty-aware inference and acquiring evidence only when needed, you can significantly improve schema completeness, relevance, and reduce token costs, as demonstrated by EviLink's 90.15% recall on Spider2-Snow. This will enhance your downstream SQL generation performance.

Key insights

EviLink reframes Text-to-SQL schema linking as uncertainty-aware inference over multiple SQL paths, improving efficiency and accuracy.

Principles

• Schema linking benefits from multi-path consideration.
• Uncertainty-guided evidence acquisition optimizes cost.
• Balancing completeness, relevance, and token cost is key.

Method

EviLink combines multi-hypothesis schema grounding with uncertainty-guided evidence acquisition to infer schema needs across multiple plausible SQL paths, distinguishing certain from uncertain items.

In practice

• Implement multi-hypothesis schema grounding.
• Integrate uncertainty-guided evidence acquisition.
• Evaluate schema linking by completeness, relevance, token cost.

Topics

• Text-to-SQL
• Schema Linking
• EviLink
• Uncertainty-Guided Evidence
• Multi-Path Grounding
• Large-Scale Databases

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• IndicDB -- Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages — IndicDB reveals a significant "Indic Gap" in LLM Text-to-SQL performance for Indian languages compared to English.
• TRACES: Tagging Reasoning Steps for Adaptive Cost-Efficient Early-Stopping — TRACES uses real-time reasoning step tagging to enable adaptive, cost-efficient early stopping in Language Reasoning Models.
• GliNER2: Extracting Structured Information from Text — GliNER2 unifies multiple NLP extraction tasks into a single, CPU-efficient, schema-driven framework.
• MARCH: Evaluating the Intersection of Ambiguity Interpretation and Multi-hop Inference — Multi-hop QA requires models to navigate layered ambiguity across complex reasoning paths.
• Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context — Self-reflection using intrinsic uncertainty signals significantly improves long-context reasoning in language models over explicit recursion.
• Why NL-to-SQL Systems Fail: Lessons from building a Semantic SQL validation pipeline — Semantic correctness and robust validation are paramount for reliable NL-to-SQL systems, beyond mere syntax.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.