Title: Reframing Hallucination Detection as Out-of-Distribution Detection: A Geometric Perspective

Abstract:

Addressing hallucinations in large language models (LLMs) remains a pivotal challenge with profound consequences for system safety and reliability. Although current detection methods demonstrate robust performance in question-answering scenarios, they often struggle with tasks that demand complex reasoning. This study reinterprets hallucination detection through the framework of out-of-distribution (OOD) detection, a thoroughly investigated domain within computer vision. By conceptualizing the next-token prediction mechanism in LLMs as a classification problem, we can leverage OOD methodologies, assuming necessary adjustments are implemented to accommodate the unique structural characteristics of language models. Our findings indicate that these OOD-inspired strategies enable the creation of detectors that are both training-free and capable of operating on single samples, delivering high accuracy in identifying hallucinations during reasoning processes. Consequently, this work posits that shifting the perspective of hallucination detection to align with OOD detection offers a viable, scalable, and promising route for enhancing the safety of language models.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC