Large Language Model Hallucination Detection and Knowledge Verification

The field of large language models (LLMs) is moving towards addressing the issue of hallucinations, which can undermine their reliability in practical applications. Researchers are exploring innovative methods to detect and verify the accuracy of LLM-generated responses, including the use of multiple small language models and integrated information theory. These approaches aim to provide a more robust and efficient solution for hallucination detection and knowledge verification. Noteworthy papers in this area include: The Trilemma of Truth in Large Language Models, which introduces a probing method called sAwMIL to assess the veracity of LLMs' internal knowledge. TUM-MiKaNi at SemEval-2025 Task 3, which proposes a multilingual hallucination identifier that combines retrieval-based fact verification with a BERT-based system to identify common hallucination patterns.

Sources

Hallucination Detection with Small Language Models

Can "consciousness" be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis

The Trilemma of Truth in Large Language Models

TUM-MiKaNi at SemEval-2025 Task 3: Towards Multilingual and Knowledge-Aware Non-factual Hallucination Identification

Built with on top of