Researchers Propose Internal Embodiment For AI

UCLA Health researchers published a Neuron paper (Akila Kadambi et al.) arguing current multimodal LLMs lack "internal embodiment," a persistent monitoring of internal states like fatigue, uncertainty, or processing load. They demonstrate measurable failures — for example, models misclassifying point-light human motion — and propose a dual-embodiment framework plus new benchmarks to improve model safety, consistency, and alignment.