Consistency-based Black-box Uncertainty Quantification for Text-to-SQL by Similarity Aggregation

Debarun Bhattacharjya, Balaji Ganesan, Radu Marinescu, Katsiaryna Mirylenka, Junkyu Lee, Michael Glass, Xiao Shou

Keywords: text-to-sql, uncertainty quantification, black-box, similarity, large language model, generative model

TL;DR: We propose and investigate black-box uncertainty quantification approaches for estimating the confidence of LLM generations in text2SQL applications by aggregating pairwise similarities between generations.

Abstract: When does a large language model (LLM) know what it does not know? Uncertainty quantification (UQ) provides an estimate of the confidence in an LLM's generated output and is therefore increasingly recognized as a crucial component of trusted AI systems. UQ is particularly important for complex generative tasks such as \emph{text-to-SQL}, where an LLM helps users gain insights about data stored in noisy and large databases by translating their natural language queries to structured query language (SQL). \emph{Black-box} UQ methods do not require access to internal model information from the generating LLM, and therefore have numerous real-world advantages, such as robustness to system changes, adaptability to choice of LLM (including those with commercialized APIs), reduced costs, and substantial computational tractability. In this paper, we investigate the effectiveness of black-box UQ techniques for text-to-SQL, where the consistency between a generated output and other sampled generations is used as a proxy for estimating its confidence. We propose a high-level non-verbalized \emph{similarity aggregation} approach that is suitable for complex generative tasks, including specific techniques that train confidence estimation models using small training sets. Through an extensive empirical study over various text-to-SQL datasets and models, we provide recommendations for the choice of sampling technique and similarity metric. The experiments demonstrate that our proposed similarity aggregation techniques result in better calibrated confidence estimates as compared to the closest baselines, but also highlight how there is room for improvement on downstream tasks such as selective generation.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11952