Neural Probe-Based Hallucination Detection for Large Language Models

TL;DR

This paper introduces a neural network-based token-level hallucination detection method for large language models, leveraging lightweight MLP probes on hidden states to improve accuracy and stability over existing techniques.

Contribution

It proposes a nonlinear probe framework with a multi-objective loss and Bayesian optimization for optimal probe placement, advancing hallucination detection in LLMs.

Findings

MLP probes outperform state-of-the-art methods in accuracy and recall.

The approach achieves low false-positive detection in multiple datasets.

Layer optimization enhances detection performance.

Abstract

Large language models(LLMs) excel at text generation and knowledge question-answering tasks, but they are prone to generating hallucinated content, severely limiting their application in high-risk domains. Current hallucination detection methods based on uncertainty estimation and external knowledge retrieval suffer from the limitation that they still produce erroneous content at high confidence levels and rely heavily on retrieval efficiency and knowledge coverage. In contrast, probe methods that leverage the model's hidden-layer states offer real-time and lightweight advantages. However, traditional linear probes struggle to capture nonlinear structures in deep semantic spaces.To overcome these limitations, we propose a neural network-based framework for token-level hallucination detection. By freezing language model parameters, we employ lightweight MLP probes to perform nonlinear modeling of high-level hidden states. A multi-objective joint loss function is designed to enhance detection stability and semantic disambiguity. Additionally, we establish a layer position-probe performance response model, using Bayesian optimization to automatically search for optimal probe insertion layers and achieve superior training results.Experimental results on LongFact, HealthBench, and TriviaQA demonstrate that MLP probes significantly outperform state-of-the-art methods in accuracy, recall, and detection capability under low false-positive conditions.