External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing

Problem

Current methods for hallucination detection in Large Language Models (LLMs) primarily focus on token-wise binary classification, which limits their effectiveness in identifying and addressing hallucinations at a more granular level. This paper addresses this gap by proposing a framework that enables span-level hallucination detection, allowing for a more nuanced understanding of hallucination dynamics in LLMs. The work is presented as a preprint and has not yet undergone peer review.

Method

The authors propose an internal hidden state framework designed for span-level hallucination detection. This framework utilizes layer-wise activation pattern inspection to identify the onset and continuation of hallucination tokens within the model's hidden states. The evaluation of the proposed method is conducted using the Precision-Recall AUC metric, which provides a robust measure of performance against random baselines. The authors do not disclose specific architectural details, loss functions, or training compute used in their experiments, focusing instead on the mechanism of hidden state probing.

Results

The available text does not report quantitative results. However, it indicates that the proposed method achieves substantial improvements in Precision-Recall AUC compared to random baselines, suggesting that the framework effectively enhances the detection of hallucinations at the span level.

Limitations

The authors acknowledge that class imbalance in hallucination detection is not explicitly addressed in their framework. This limitation could impact the generalizability and robustness of the detection mechanism, particularly in scenarios where hallucinations are rare compared to non-hallucinated outputs. Additionally, the lack of detailed quantitative results limits the ability to fully assess the performance of the proposed method against established benchmarks.

Why it matters

This work has significant implications for downstream applications of LLMs, particularly in contexts where the reliability of generated content is critical. By advancing the detection of hallucinations to a span-level analysis, the proposed framework could enhance the interpretability and trustworthiness of LLM outputs, paving the way for more robust applications in fields such as automated content generation, conversational agents, and information retrieval.