Integrating Machine Consciousness Simulation and LLMs Toward Verbal and Non-Verbal General Intelligence in Artificial Agents
Résumé
Recent advances in artificial general intelligence (AGI) reveal an asymmetry: large language models (LLMs) demonstrate human-like linguistic reasoning, yet they may fall short in non-verbal reasoning and embodied cognition. To bridge this gap, we propose a computational framework unifying verbal and non-verbal intelligence through consciousness-inspired modeling. Leading theories of consciousness, such as Global Workspace Theory (GWT) and Integrated Information Theory (IIT), remain inadequate due to imprecise formalism and computational intractability. In contrast, the Projective Consciousness Model (PCM)-a tractable geometrical framework modeling spatial perspective, emotion, and attention-can be integrated with LLMs. The resulting PCM-LLM architecture fuses belief states in projective space with linguistic representations, enabling coherent verbal and non-verbal reasoning and belief updating. In a competitive social task requiring multimodal integration, PCM-LLM agents outperformed PCM-only agents, highlighting language's broader role in cognitive dynamics, beyond its expressive functions. Such integration of human-inspired consciousness models and LLMs may help guide the development of interpretable, human-like AGI.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |