Overview
This report details a conceptual unified framework for world models, aiming to incorporate cognitive functions observed in human cognition. It distinguishes existing world model research based on the specific cognitive functions they innovate, asserting a need for grounding in principles from human and machine cognition theory to evaluate claims of human-like capabilities.
Research Context
The research addresses the observed trend of various works claiming human-like cognitive capabilities in their world models. The framework is presented as a means to move towards models that exhibit a broader range of such capabilities. Prior taxonomies are noted as not suggesting the same research directions identified by this report's taxonomy.
Approach
The approach involves developing a conceptual unified framework that comprehensively integrates a set of cognitive functions. These functions include memory, perception, language, reasoning, imagining, motivation, and metacognition. The framework is then used to identify gaps in existing research within the field of world models.
Findings
- The report identifies that motivation, specifically intrinsic motivation, and metacognition are significantly under-researched areas within current world model developments.
- Concrete directions are proposed to address these gaps, informed by active inference and global workspace theory.
- A new category, "epistemic world models," is introduced. This category encompasses agent frameworks designed for scientific discovery, operating over structured knowledge.
- The taxonomy developed within this framework, when applied to video, embodied, and epistemic world models, suggests research directions that were not indicated by previous taxonomies.
- The framework incorporates all listed cognitive functions: memory, perception, language, reasoning, imagining, motivation, and metacognition.
Why This Matters
The proposed framework provides a structured approach to evaluate and guide the development of world models towards more human-like cognitive capabilities. By identifying specific under-researched areas and introducing new conceptual categories, it offers a guide for future states of the art in the field of artificial intelligence.