Leveraging Graph Structures to Detect Hallucinations in Large Language Models

TL;DR

This work addresses hallucinations in large language models by proposing a graph-based detection approach that leverages latent-space structure and homophily. It constructs a semantic similarity graph from sentence embeddings and trains a Graph Attention Network to classify generations as hallucinated or true, with a label-recovery task for new data and contrastive learning to boost robustness. The method achieves competitive results on self-generated data and benchmark datasets (FEVER and SelfCheckGPT) without relying on external knowledge, and shows the benefit of non-local aggregation and graph-based message passing for hallucination detection. The findings highlight the latent-space structure of hallucinations and offer a scalable, model-agnostic framework that could generalize to other labeling tasks beyond hallucination detection.

Abstract

Large language models are extensively applied across a wide range of tasks, such as customer support, content creation, educational tutoring, and providing financial guidance. However, a well-known drawback is their predisposition to generate hallucinations. This damages the trustworthiness of the information these models provide, impacting decision-making and user confidence. We propose a method to detect hallucinations by looking at the structure of the latent space and finding associations within hallucinated and non-hallucinated generations. We create a graph structure that connects generations that lie closely in the embedding space. Moreover, we employ a Graph Attention Network which utilizes message passing to aggregate information from neighboring nodes and assigns varying degrees of importance to each neighbor based on their relevance. Our findings show that 1) there exists a structure in the latent space that differentiates between hallucinated and non-hallucinated generations, 2) Graph Attention Networks can learn this structure and generalize it to unseen generations, and 3) the robustness of our method is enhanced when incorporating contrastive learning. When evaluated against evidence-based benchmarks, our model performs similarly without access to search-based methods.

Figures (3)

• Figure 1: Data generation process.
• Figure 2: Prompt 1: Without provided Context
• Figure 3: Prompt 2: With provided Context