← News

ResearchSources: arXiv · 18 Nov 2026

DisGuAD uses disagreement among candidate SQL queries to pick which ambiguity to clarify

A research team proposes a framework that treats differing plausible SQL interpretations as evidence of which user question matters most, reporting gains on BIRD-Interact benchmarks.

In plain words

Text-to-SQL systems turn plain-language questions into database queries, and when a request is ambiguous the agent must spend its limited follow-up questions wisely. DisGuAD generates several possible SQL interpretations, compares where they differ, and asks the user about the difference that most affects the final query. The authors report higher task success and better clarification choices than the baseline policy.

What was announced

The paper, an arXiv abstract by the DisGuAD authors, argues that current clarification policies have three weaknesses: they flag surface ambiguity without checking its effect on the SQL, they miss ambiguities hidden in schema routes or domain knowledge, and they waste limited interaction turns on unfocused or redundant questions. The core challenge, it says, is identifying ambiguity that changes the SQL before the intended query is known.

DisGuAD builds an inventory of query terms, elicits contrastive SQL interpretations through complementary probe families, compares their canonical SQL components, and attributes recurring differences to user-facing clarification targets; a lightweight prioritization and dialogue-state mechanism then selects one high-impact unresolved target. On BIRD-Interact Lite and Full with two LLM backbones, the authors report normalized Reward improvements of 4.40 to 8.97 points over the original interaction policy, critical-ambiguity coverage rising from 48.62% to 82.75% and from 45.76% to 85.14%, and gains of up to 13.97 Reward points when its clarification signals are added to two static text-to-SQL frameworks with their SQL generators unchanged. These figures come from the authors' own experiments and have not been independently verified.

What it does not show

  • The evidence is the paper's abstract only; experimental setup, baselines and failure cases cannot be assessed from this excerpt.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.