SiGHT: A Self-Supervised Graph-based Hallucination DeTection Framework for Domain-Specific LLMs

Zi-Ying Chen, Meng-Fen Chiang, Wen-Chih Peng

TL;DR: We introduce SiGHT, a novel framework that combines synthetic hallucination generation and graph-based learning for robust hallucination detection in domain-specific LLMs.

Abstract: Hallucination detection is vital for early error diagnosis in domain-specific Large Language Models (LLMs). In this paper, we propose SiGHT, a novel framework for hallucination detection tailored to domain-specific LLMs. Unlike existing approaches that rely heavily on costly retrieval-based mechanisms or manually annotated datasets, the proposed framework introduces a fully automatic training data generation that uses prompt-based strategies to synthesize hallucinated content from structured domain knowledge, thus eliminating the need for human annotation. Specifically, it transforms both original knowledge data and generated hallucinated data into word-level graphs, which are then processed using a Graph Attention Network (GAT), which applies attention mechanisms over the graph structure to model fine-grained relational dependencies. On MSMARCO-QnA and RAGTruth-QA, SiGHT yields a 37.31\% relative F1 gain over prior graph-based baselines and achieves competitive performance with state-of-the-art detectors, while using only 0.03M parameters and incurring 0.342 s inference latency. By combining automatic data generation with an efficient graph encoder, SiGHT offers favorable accuracy–efficiency trade-offs and eliminates human annotation, making it practical for deployment in domain-specific pipelines.

Submission Number: 1135