Proceedings of the Annual Meeting of the Cognitive Science Society

Personality computing has long struggled to bridge the gap between high-level psychological theories and low-level computational features. To address the lack of psychological grounding in personality computing, we propose a multimodal framework based on the Cognitive-Affective Personality System (CAPS). We decompose inputs into "Cognitive Units" (semantic construal) and "Affective Units" (emotional reactivity). Specifically, we introduce a parameter-free Matrix-Based Entropy mechanism to dynamically weight hidden states from Qwen2.5-3B, treating representational entropy as a measure of cognitive complexity. For audio, we combine linguistic content (Wav2Vec2) with prosodic features (OpenSMILE). Evaluated on a new self-reported Chinese dataset, our approach significantly outperforms unimodal baselines. Crucially, interpretability analysis reveals that the model's internal entropy profiles spontaneously recover latent psychological meta-traits (Plasticity and Stability), demonstrating that our architecture effectively bridges deep learning features with the intrinsic structure of human personality.