World Models and Causal Reasoning in General Intelligence Systems: A Pathway to Human-Level Decision-Making
Abstract
Artificial General Intelligence (AGI) requires systems capable of robust reasoning, out-of-distribution (OOD) generalization, and counterfactual foresight. Contemporary deep learning models, while exhibiting unprecedented capabilities in pattern recognition and predictive text generation, remain fundamentally constrained by correlation-based statistical inference. These association-based approaches struggle with structural interventions, physical common-sense reasoning, and counterfactual evaluation—key pillars of human cognitive architecture. This comprehensive review paper systematically explores the convergence of internal World Models and Causal Reasoning frameworks as a viable computational pathway toward human-level decision-making in AGI systems. We analyze how generative world models synthesize internal representations of environment dynamics and evaluate how Structural Causal Models (SCMs), interventional calculus (do-calculus), and counterfactual inference can be embedded into latent generative states. Through a formal comparative framework, we evaluate existing paradigms across reinforcement learning, model-based control, and generative AI. Furthermore, we outline key research gaps in neuro-symbolic integration, real-time causal discovery, and scalability. Finally, we propose a novel conceptual taxonomy and hybrid architecture combining structural causal discovery with latent generative world models, charting future research trajectories for autonomous general intelligence systems.
KEYWORDS: Artificial General Intelligence, World Models, Causal Reasoning, Structural Causal Models, Counterfactual Inference, Reinforcement Learning, Out-of-Distribution Generalization.
Full Text:
PDF 123-135Refbacks
- There are currently no refbacks.