[2E4-OS-4a] OS-4

[2E4-OS-4a-04] Toward Safer World and Agent Modeling with Natural Semantic Metalanguage-based Controlled Cognition

〇RAFAL RZEPKA¹ (1. Hokkaido University)

Keywords:Knowledge Acquisition, Cognition Simulation, Transparency, Explainability, Interpretability

The development of foundation models has shown how large datasets and powerful computing can create useful tools for daily life. However, these models lack agency and fail to ground language in sensory experience. Current methods for building intelligent entities rely on pre-training with massive datasets and using human annotators to teach models safe behavior. This approach is inefficient, unsustainable, and its safety remains uncertain. In this paper, I argue for a shift toward knowledge acquisition methods based on readable cognitive concepts rather than opaque weight-based representations. As the autonomy of physical agents grows with technological advancements, higher-level processing is needed to enable transparent tracking of how an agent’s behavior is designed. I hypothesize that a fixed set of semantic building blocks for perception and learning could improve the explainability of artificial entities. The Natural Semantic Metalanguage framework offers a promising example of how such a set of basic perceptual concepts might be defined.