Proceedings of the Annual Meeting of the Cognitive Science Society

Current multimodal visual models continue to improve in perceptual performance, yet still rely on passive, dense processing, resulting in high computational overhead. From a cognitive perspective, existing methods incorporate limited prior memory and resource allocation during perception, making it difficult to construct perception as goal-directed sequential decision-making. This paper proposes an active metacognitive framework inspired by biological memory priors, organizing visual understanding as a memory-driven hypothesis–verification process. The model forms initial beliefs from low-resolution global information and internal memory, and selects local observations to reduce uncertainty. As evidence accumulates, the system adaptively terminates perception based on confidence, enabling on-demand allocation of computational resources. Experiments across visual complexity settings show that this approach matches or surpasses traditional models while reducing inference time and token usage by approximately 15–30%, and produces structured, interpretable observation sequences.