Uncertainty Quantification with Generative-Semantic Entropy Estimation for Large Language Models
Keywords: Large Language Models, Uncertainty Quantification, Explainable AI, Trustworthy AI
TL;DR: We introduce Generative-Semantic Entropy Estimation (GSEE), a lightweight, model-agnostic algorithm to estimate uncertainty quantification for foundational and related models by estimating the entropy of the subspace spanned by model responses.
Abstract: In recent years, powerful foundation models, including Large Language Models (LLMs) and Large Multi-Modal Models (LMMs) have ushered in a new epoch of multi-faceted, intelligent conversational agents. Despite their significant early successes and widespread use, foundation models nevertheless currently suffer from several critical challenges, including their lack of transparency and predilection for "hallucinations." To this end, we introduce Generative-Semantic Entropy Estimation (GSEE), a model-agnostic algorithm that efficiently estimates the generative uncertainty associated with foundation models, while requiring no additional auxiliary model inference steps. In principle, for any foundation model input data, e.g., a text prompt, image, text + image, etc., GSEE numerically estimates the uncertainty encapsulated in the internal, semantic manifold of the LLM generated responses to the input data. In this way, high uncertainty is indicative of hallucinations and low generative confidence. Through experiments, we demonstrate the superior performance of GSEE for uncertainty quantification (UQ) amongst state-of-the-art methods across a variety of models, datasets, and problem settings, including: unbounded language prompting, constrained language prompting, high/low generative stochasticity, acute semantic diversity prompting, and as a barometer for hallucination/predictive accuracy.
Primary Area: interpretability and explainable AI
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.
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: 5682
Loading