Multimodal models and world models are emerging as promising frameworks for extending language-based AI beyond text, towards systems that can represent, predict, and interact with aspects of the physical world. In doing so, they appear to move closer to forms of embodied cognition observed in biological agents. However, building such systems requires more than adding sensory modalities to existing language models. It requires confronting the structural, physical, and dynamical constraints that shape how information is represented, integrated, remembered, and transformed over time.

A growing body of work suggests that multimodal systems may remain cognitively limited if their internal representational spaces are fixed, externally imposed, or only weakly shaped by their own activity. Without mechanisms for endogenous feedback, self-organisation, and representational restructuring, such systems may be capable of sophisticated pattern integration while still lacking the capacity for genuine cognitive development. Even when debates about grounding and embodiment are set aside, important questions remain about how multimodal models internalise experience, update their world models, and reorganise their own representational hierarchies.

This Research Topic asks how architectural design, physical dynamics, memory systems, and representational flexibility might allow multimodal world models to function not only as more capable AI systems, but also as experimental tools for probing the conditions of human-like cognition.

This Research Topic aims to identify the physical, architectural, and functional principles that could allow multimodal systems to exhibit genuine cognitive amplification, rather than surface-level linguistic mimicry. A central question is whether current architectures possess the structural plasticity, dynamic memory, and temporal depth required for cognitive growth through language and reasoning - and if not, why.

The Topic brings together perspectives from cognitive science, artificial intelligence, neuroscience, philosophy of mind, and complex systems theory. Its purpose is to clarify the conditions under which multimodal world models might reshape their own cognitive spaces through interaction and feedback.

Guiding questions

We welcome contributions addressing, but not limited to, the following questions:

• What architectural, dynamical, or physical constraints shape the cognitive capacities of multimodal systems, and how do these constraints limit development beyond language simulation?
• How should the relationship between grounding, embodiment, multimodality, and sensorimotor coupling be understood in artificial cognitive systems?
• What forms of cognitive development, if any, can emerge from language-only systems, and how do these differ from those enabled by multimodal or embodied interaction?
• Under what conditions can artificial agents reorganise their representational hierarchies, develop metacognitive functions, and expand their cognitive repertoire over time?
• In what ways can language amplify, redirect, or transform cognition, and what roles do transformative memory, temporal depth, and path-dependent dynamics play in this process?
• How might neuromorphic, hybrid, embodied, or self-modifying architectures change the limits of current AI systems and address forms of computational closure?
• How can artificial world models be used to test, refine, or challenge theories of human cognition, including theories of embodiment, abstraction, self-organisation, and cognitive transformation?

We welcome theoretical, computational, empirical, and interdisciplinary work at the intersection of language, embodiment, multimodal representation, artificial intelligence, and cognition. Contributions are especially encouraged on the following themes.

Structural and physical limits of multimodal models

• Architectural closure and restricted representational state spaces
• Static versus dynamically reconfigurable representational structures
• Limits of current memory architectures for path-dependent cognitive development
• Absence of self-organising, meta-structural, or adaptive mechanisms
• Structural atemporality and its implications for cognition
• Temporal coupling, feedback, and recurrence as conditions for cognitive reframing
• Differences between pattern integration, world modelling, and cognitive transformation

Language as a cognitive amplifier
• Language as a driver of abstraction, generalisation, planning, and representational expansion
• Interactions between linguistic representation, intentionality, self-reflection, and metacognition
• The role of language in reorganising perceptual, motor, and conceptual spaces
• Transformative memory as a condition for learning from linguistic and social interaction
• Whether language alone can support cognitive growth, or whether it requires embodiment and feedback-rich dynamics

Beyond current architectures
• Neuromorphic, analogue, and non-von Neumann computation
• Hybrid memory systems, including memristive, hysteretic, adaptive, and reconstructive forms
• Dynamically reconfigurable and open-ended hardware architectures
• Embodied and sensorimotor architectures capable of adaptive world modelling
• Artificial systems with self-modifying, recursive, or developmentally open structures
• Complex systems approaches to cognition, self-organisation, and emergent agency

Are you interested in contributing a manuscript to this Research Topic?

Multimodal World Models, Embodiment, and Cognitive Amplification