Scalable and Enhanced Hallucination Detection in LLMs using Semantic Clustering

Samir Abdaljalil, Parichit Sharma, HASAN KURBAN, Rachad Atat, Erchin Serpedin, Khalid Qaraqe

Keywords: Hallucination, Semantic entropy, LLMs, Semantic clustering

Abstract: Large language models (LLMs) are increasingly being adopted across various domains, driven by their ability to generate general-purpose and domain-specific text. However, LLMs can also produce responses that seem plausible but are factually incorrect—a phenomenon commonly referred to as "hallucination". This issue limits the potential and trustworthiness of LLMs, especially in critical fields such as medicine and law. Among the strategies proposed to address this problem uncertainty-based methods stand out due to their ease of implementation, independence from external data sources, and compatibility with standard LLMs. In this paper, we present an optimized semantic clustering framework for automated hallucination detection in LLMs, using sentence embeddings and hierarchical clustering. Our proposed method enhances both scalability and performance compared to existing approaches across different LLM models. This results in more homogeneous clusters, improved entropy scores, and a more accurate reflection of detected hallucinations. Our approach significantly boosts accuracy on widely used open and closed-book question-answering datasets such as TriviaQA, NQ, SQuAD, and BioASQ, achieving AUROC score improvements of up to 9.3% over the current state-of-the-art semantic entropy method. Further ablation studies highlight the effectiveness of different components of our approach.

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: 10575