Embodiment in Multimodal Large Language Models

Current Multimodal Large Language Models (MLLMs) demonstrate an extraordinary capacity to bridge textual and visual inputs, yet they face significant limitations in situated physical and social interactions.

These models lack bodily experience, understanding the world largely through statistical associations. To achieve greater world comprehension, we posit that models must move beyond token associations toward "felt" cognition supported by internal and external models.

This workshop introduces our Dual-Embodiment Framework proposed recently in Neuron (Kadambi et al., 2026), which integrates two complementary domains: external embodiment (interactions with the external world) and internal embodiment (internal dynamics, proprioception, and homeostasis). We will explore how explicitly operationalizing internal embodiment through our definition of continuous state modeling can be applied toward improvements in AI performance and safety.

Our goal is to convene researchers across frontier disciplines in AI, cognitive science, neuroscience, and robotics, to discuss translating these concepts into implementable architectures and define new internally-aligned benchmarks.

Following the workshop, selected speakers, panelists, and interested attendees will collaborate to draft the definitive Vision & Roadmap Paper on Internal Embodiment in AI.

Rather than a traditional call for papers, our workshop focuses on interactive dialogue, synthesis, and collaborative community output. Attendees interested in co-authoring and shaping the future research agenda will join our post-workshop working group across three core pillars: