Do small and medium-sized LLMs simulate embodied cognition? A preliminary analysis for enhancing ITS with embodied cognition capabilities

Large Language Models (LLMs) lack embodied cognition, as theyoperate through symbolic processing without grounding in sensoryor motor experience, thus facing the symbol grounding problem.Their knowledge derives from external data rather than from livedinteraction, preventing access to situational intelligence. This limi-tation also affects Intelligent Tutoring Systems (ITS) in domains thatrequire embodied skills, where LLMs cannot autonomously con-struct scenarios or evaluate learners’ actions. This work presentsan initial analysis of how small and medium-sized LLMs can simu-late aspects of embodied cognition through specifically designedprompts. A dataset inspired by Gibson’s theory of affordances anda prompt template based on the concept of proxied experience weredeveloped. Experiments on a selected set of LLMs yielded promis-ing results, suggesting potential applications in training contextsthat rely on embodied learning