New AI Research Framework Proposes Unified World Models with Cognitive Architecture Theory
A detailed research document presents a unified conceptual framework for world models in artificial intelligence, based on Cognitive Architecture Theory (CAT). This framework integrates all cognitive functions related to CAT, such as memory, perception, language, reasoning, imagination, motivation, and meta-cognition. The report highlights the uniqueness of previous studies by the cognitive functions they enhance and assesses claims regarding "human-like" cognitive abilities in current world models. It points out critical research deficiencies, particularly in intrinsic motivation and meta-cognition, which are significantly overlooked. To fill these gaps, the authors suggest specific research paths inspired by active inference and global workspace theory. Additionally, the paper introduces a new category termed Epistemic World Models, focusing on agent frameworks for scientific discovery. This research aims to guide future advancements in AI world models. The report is available on arXiv under the identifier arXiv:2604.16592v1 and is classified as a cross-disciplinary announcement.
Key facts
- The report presents a unified framework for world models based on Cognitive Architecture Theory (CAT).
- It evaluates claims of "human-like" cognitive capability in existing AI world models.
- The framework incorporates seven cognitive functions: memory, perception, language, reasoning, imagining, motivation, and meta-cognition.
- Motivation (especially intrinsic motivation) and meta-cognition are identified as drastically under-researched areas.
- Concrete research directions are proposed using active inference and global workspace theory.
- A new category called Epistemic World Models is introduced for agent frameworks in scientific discovery.
- The report distinguishes prior works by the cognitive functions they innovate.
- It was published on arXiv with the identifier arXiv:2604.16592v1 as a cross-disciplinary announcement.
Entities
Institutions
- arXiv