AI
Geometric Framework Uses Embedding Space to Spot and Fix AI Hallucinations
Researchers propose a new method that models semantic distributions to assess language model reliability at both prompt and answer levels.
Large language models frequently produce fluent-sounding but factually incorrect responses—a problem known as hallucination. While uncertainty quantification has been explored as a potential solution, prior approaches have lacked a coherent way to evaluate reliability across both questions and individual answers. A new geometric framework addresses this gap by treating the space of model outputs as a mathematical object to be analyzed.
The Approach#
The method works by having a language model generate multiple candidate answers for each prompt. Rather than treating these outputs independently, the researchers use archetypal analysis—a statistical technique—to map the answers onto an embedding space and estimate their collective geometric structure. This yields a model of the underlying semantic distribution without requiring access to the model's internal weights or outputs, making it agnostic to the specific architecture used.
Two complementary metrics then emerge from this geometric picture. At the prompt level, the framework approximates the entropy of the distribution to gauge how uncertain the model is about its answer. For individual answers, it measures atypicality—how far each response sits from the center of the batch—to score its reliability. The most reliable answer can then be selected directly.
Testing and Results#
Experiments on short-form question-answering benchmarks show the method performs on par with or surpasses existing uncertainty detection techniques. More importantly, it achieves notably stronger results on medical datasets, where the stakes of hallucinations are highest and accuracy is paramount.
Why It Matters#
The authors argue that thinking of language model uncertainty through the lens of semantic distributions offers both practical and theoretical benefits. It provides a principled, black-box approach that works with any model capable of generating text, while grounding the analysis in geometric intuition rather than heuristics alone.
