The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans

Researchers studying 21 large language models found a significant 'grounding gap' in how LLMs understand abstract concepts compared to humans. While LLMs rely heavily on word associations, they systematically underreproduce emotional and internal-state properties, achieving maximum correlation of r=0.37 versus human-to-human baselines above r=0.9. The findings suggest current models can identify grounding dimensions when explicitly queried but fail to recruit them naturally during free generation.

This research addresses a fundamental limitation in how large language models process meaning. The grounding gap reveals that LLMs lack the embodied, emotionally-informed understanding humans develop through lived experience. When generating text about abstract concepts like justice or theory, models default to statistical word patterns rather than integrating sensorimotor, emotional, and social dimensions that anchor human cognition. This gap persists across frontier and open-weight models, indicating it's not a limitation of specific architectures but a structural challenge in language-only learning. The finding that larger models show better alignment on explicit rating tasks suggests scaling alone doesn't solve the problem of naturally recruiting grounding information. The sparse autoencoder analysis demonstrates that relevant features exist within model internals—the issue is deployment during generation, not availability. For AI development, this indicates that scaling to human-level performance on abstract reasoning may require architectures that integrate multimodal inputs or explicit mechanisms to weight emotional and social context. The research has implications for applications requiring nuanced understanding of abstract concepts, from educational tools to legal and philosophical reasoning systems. As the field pushes toward more general AI systems, addressing grounding gaps becomes increasingly important for developing models that understand meaning the way humans do, rather than merely mimicking statistical patterns. Current deployment practices may be sufficient for many use cases, but limitations emerge in domains requiring deep conceptual reasoning.

• →LLMs show correlation of only r=0.37 with human responses on abstract concept properties, far below human-to-human baselines of r>0.9.
• →Models over-rely on word associations while systematically under-producing emotional and internal-state properties compared to human reasoning.
• →Larger models align better with human judgment on explicit grounding tasks, but this improvement doesn't translate to natural generation.
• →Sparse autoencoders reveal that grounding-relevant features exist in model internals but aren't recruited during free-text generation.
• →Addressing grounding gaps may require architectural changes beyond scaling, such as multimodal integration or explicit context-weighting mechanisms.