Proceedings of the Annual Meeting of the Cognitive Science Society

Psychological defense mechanisms serve as unconscious cognitive strategies aimed at regulating anxiety and protecting the self. These processes are not limited to clinical settings but are also prevalent in social media interactions. While Large Language Models (LLMs) can capture semantic ambiguity, a structural misalignment arises as their decoding patterns yield spiky probability distributions with high confidence, failing to reflect the uncertainty in the spectral nature of human cognition. To bridge this gap, we propose DefenseAlign for psychological defense mechanism classification, aligning LLM confidence from pre-softmax logits with human uncertainty represented by a human belief distribution. DefenseAlign pools logits across label verbalizations, applies learnable temperature scaling to fit the human belief distribution, and distills the aligned distributions into a student model. We evaluate the framework on DefenseSpectrum, a dataset that estimates human uncertainty directly from annotation frequencies. Experiments show reduced divergence and improved preservation of secondary mechanisms.