[1Win4] Poster session 1

[1Win4-29] Proposal of a Hallucination Detection Method Using Counterargument and Reevaluation Prompts

〇Asuka Yamazato¹, Kohei Koyama¹ (1.ARISE analytics. inc)

Keywords:Large Lunguage Model, Hallucination, Prompt Engineering

Since the release of ChatGPT, generative AI adoption has advanced significantly. However, a major challenge is hallucination, where models generate inaccurate information. Conventional techniques, such as retrieval-augmented generation, require external information preparation, limiting their adaptability. Another approach, SelfCheckGPT, detects hallucination by analyzing response similarity from same multiple prompts but demands a large number of tokens, increasing costs. This study proposes a method to detect hallucination without external information and with reduced token usage. Based on the hypothesis that large language models (LLMs) reason like humans, our approach encourages counterarguments and reevaluation during reasoning to detect hallucination. Experiments using a Japanese general knowledge dataset showed that our method achieved higher Recall while reducing token usage compared to existing methods, demonstrating its potential for efficient and accurate hallucination detection.