CHAIR -- Classifier of Hallucination as Improver

TL;DR

CHAIR is a supervised framework that detects hallucinations in language models by analyzing internal logits, significantly improving detection accuracy and offering insights for better decoding strategies.

Contribution

The paper introduces CHAIR, a novel method using internal logits for hallucination detection, enhancing accuracy and robustness in zero-shot scenarios.

Findings

CHAIR achieves high detection accuracy on TruthfulQA and MMLU datasets.

It demonstrates robustness and generalizability across different datasets.

The approach enables potential improvements in decoding strategies to reduce hallucinations.

Abstract

In this work, we introduce CHAIR (Classifier of Hallucination As ImproveR), a supervised framework for detecting hallucinations by analyzing internal logits from each layer of every token. Our method extracts a compact set of features such as maximum, minimum, mean, standard deviation, and slope-from the token logits across all layers, enabling effective hallucination detection without overfitting. Experiments on TruthfulQA and MMLU datasets demonstrate that CHAIR significantly improves detection accuracy, particularly in zero-shot scenarios, showcasing its robustness and generalizability. Beyond hallucination detection, CHAIR highlights the potential of using internal representations for designing advanced decoding strategies. By leveraging patterns in logits, we suggest that more sophisticated models and adaptive decoding methods could further reduce hallucinations and enhance text completion quality. CHAIR not only offers a practical solution for detecting hallucinations but also lays the groundwork for exploring richer representations in LLMs to improve their factuality and coherence.