Proceedings of the Annual Meeting of the Cognitive Science Society

The hallucination issue in large language models (LLMs) significantly restricts their application in high-stakes domains such as healthcare, cognitive science and law. Existing approaches primarily focus on data optimization or decoding strategies but lack a fine-grained analysis of the underlying mechanisms of hallucinations. This paper proposes a psychology-guided two-stage fine-grained editing and sampling framework (PGFES), which, for the first time, introduces psychological classifications of hallucinations into LLM optimization. Firstly, an attention-augmented MLP probe is designed to identify "truthfulness directions" corresponding to different hallucination types through feature channel reweighting, enabling fine-grained editing of the model's internal representations during inference. Then, a dynamic weighting mechanism based on Jaccard similarity is employed to compute the weights of multi-path edited outputs, achieving adaptive sampling. Experiments demonstrate that the optimization method incorporating psychology-related concepts improves truthfulness by 20.4% on the TruthfulQA open-domain question-answering task compared to baseline models and exhibits strong generalization across cross-domain datasets.