Large language models without grounding recover non-sensorimotor but not sensorimotor features of human concepts

Abstract

To what extent can language give rise to complex conceptual representation? Is multisensory experience essential? Recent large language models (LLMs) challenge the necessity of grounding for concept formation: whether LLMs without grounding nevertheless exhibit human-like representations. Here we compare multidimensional representations of ~4,442 lexical concepts between humans (the Glasgow Norms1, N = 829; and the Lancaster Norms2, N = 3,500) and state-of-the-art LLMs with and without visual learning, across non-sensorimotor, sensory and motor domains. We found that (1) the similarity between model and human representations decreases from non-sensorimotor to sensory domains and is minimal in motor domains, indicating a systematic divergence, and (2) models with visual learning exhibit enhanced similarity with human representations in visual-related dimensions. These results highlight the potential limitations of language in isolation for LLMs and that the integration of diverse modalities can potentially enhance alignment with human conceptual representation.