Detecting Hallucinations in Scientific Claims by Combining Prompting Strategies and Internal State Classification

Large Language Model (LLM) based research assistant tools demonstrate impressive capabilities, yet their outputs may contain hallucinations that compromise their reliability.Therefore, detecting hallucinations in automatically generated scientific content is essential.Sci-Hal2025: Hallucination Detection for Scientific Content challenge @ ACL 2025 provides a valuable platform for advancing this goal.This paper presents our solution to the Sci-Hal2025 challenge.Our approach combines several prompting strategies to prompt LLMs and leverages their hidden states as features to build the classifier.We first benchmark multiple LLMs on the SciHal dataset under the zero-shot prompting.Next, we developed a detection pipeline that integrates few-shot and chain-of-thought prompting.Then, the hidden representations extracted from the LLMs serve as features for an auxiliary classifier, further improving detection performance.In this paper, we present comprehensive experimental results and discuss the implications of our findings for future research on hallucination detection in scientific content.