Grounded world models in biological organisms and future embodied AI

Recent breakthroughs in generative AI, particularly Large Language Models (LLMs), have demonstrated that predictive learning on massive linguistic corpora can capture significant statistical regularities about the world. However, the authors argue that this approach—where language acts as the primary scaffold for all other knowledge—is fundamentally different from biological intelligence. In humans and other animals, intelligence is rooted in grounded world models developed through active, embodied interaction with the physical and social environment. These models provide the semantic basis upon which language and higher-level reasoning are later constructed.

Biological organisms do not learn from passive, curated datasets. Instead, they engage in continuous action-perception loops to satisfy homeostatic needs. The authors highlight five key neural systems that illustrate this grounded approach:

1. Navigation: The hippocampal-entorhinal system uses intrinsic, preconfigured dynamics (like grid cells) to map physical space, which are then calibrated through movement and later redeployed for conceptual navigation.
2. Affordance-based perception: Organisms perceive objects not just as visual entities, but in terms of the actions they afford (e.g., graspability).
3. Active perception: Learning is driven by exploration rather than passive observation.
4. Allostatic control: Emotional and homeostatic regulation provides a foundation for valuing outcomes.
5. Self-world distinction: The ability to differentiate between self-generated and world-generated outcomes is critical for agency.

To move beyond the limitations of current embodied AI, the authors suggest shifting toward training regimes that prioritize autonomous experience and open-ended learning. By incorporating principles such as intrinsic neural dynamics and social interaction, future AI could develop world models that are not only grounded in physical reality but also aligned with human norms and values. This transition would allow AI to move from merely predicting the next token to truly understanding the consequences of actions in a complex, dynamic world.