What do Geometric Hallucination Detection Metrics Actually Measure?

TL;DR

This paper investigates what specific aspects of hallucinations in language models are captured by geometric detection metrics, revealing their sensitivities and proposing normalization to improve multi-domain robustness.

Contribution

The study clarifies what geometric hallucination detection metrics measure and introduces a normalization method to reduce domain shift sensitivity, improving detection performance.

Findings

Different geometric statistics capture different hallucination types.

Existing metrics are sensitive to domain shifts.

Normalization improves AUROC by +34 points in multi-domain settings.

Abstract

Hallucination remains a barrier to deploying generative models in high-consequence applications. This is especially true in cases where external ground truth is not readily available to validate model outputs. This situation has motivated the study of geometric signals in the internal state of an LLM that are predictive of hallucination and require limited external knowledge. Given that there are a range of factors that can lead model output to be called a hallucination (e.g., irrelevance vs incoherence), in this paper we ask what specific properties of a hallucination these geometric statistics actually capture. To assess this, we generate a synthetic dataset which varies distinct properties of output associated with hallucination. This includes output correctness, confidence, relevance, coherence, and completeness. We find that different geometric statistics capture different types of hallucinations. Along the way we show that many existing geometric detection methods have substantial sensitivity to shifts in task domain (e.g., math questions vs. history questions). Motivated by this, we introduce a simple normalization method to mitigate the effect of domain shift on geometric statistics, leading to AUROC gains of +34 points in multi-domain settings.