A Proposed Integrated Multi-Signal Framework for Detecting Hallucinations in Large Language Model-Generated Text

The Large Language Model (LLMs) have proven to be extremely powerful in generating high-quality fluent and informative responses to a wide range of prompts. The main issue with such model is that they tend to generate hallucinated responses i.e. responses that contain information not supported by the context. The issue with detecting hallucinations is that a response might sound reasonable, even if it contains irrelevant or incorrect information. The prior work on the topic has had limited success, as the majority of it focuses on utilizing only one type of evidence to detect hallucination, which might be insufficient to capture the complexity of phenomenon. This paper proposes a novel method for detecting hallucination in LLMs responses that utilize a hybrid of different signals to detect hallucinations. This paper introduces a set of signals based on semantic consistency, semantic uncertainty, factual or contextual verification, prompt sensitivity and model variability that evaluate the responses from different angles to determine whether it contains any hallucinations. The paper tests the model on the HaluEval dataset and compares it to individual signals to see whether the hybrid approach would perform better than any single signal. It evaluates the performance using standard classification evaluation metrics, including accuracy, precision, recall, F1, ROC-AUC and PR-AUC and performs an ablation study to assess the impact of each individual signal. Overall, this paper aims to improve the current understanding of the hallucination detection task and the factors influencing it and provide an effective solution to the problem.