What do Geometric Hallucination Detection Metrics Actually Measure?
This work addresses the problem of understanding and improving hallucination detection for deploying generative models in high-consequence applications, but it is incremental as it builds on existing geometric methods.
The paper investigates what specific hallucination properties geometric detection metrics capture by generating a synthetic dataset that varies correctness, confidence, relevance, coherence, and completeness, finding that different statistics capture different types of hallucinations. It also introduces a normalization method to mitigate domain shift effects, achieving AUROC gains of +34 points in multi-domain settings.
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.