Proceedings of the Annual Meeting of the Cognitive Science Society

Large language models encode rich semantic information, but how concreteness is represented across layers remains unclear. We examine layer-wise linear separability of concreteness by training linear probes on hidden representations from two open-source model families at multiple scales: Qwen3 and Gemma3-Instruct. Using human concreteness ratings, we build balanced prompt datasets with four difficulty levels: an extreme abstract–concrete contrast and three finer boundary comparisons at the abstract end, mid-range, and concrete end. Probes achieve high accuracy on the extreme contrast in shallow layers, showing that endpoint differences are strongly linearly separable. For finer distinctions, performance follows a stable hierarchy: mid-range concreteness is easiest to separate, abstract-end distinctions are hardest, and concrete-end distinctions are intermediate. Across models, accuracy rises rapidly in early layers, peaks in middle layers, and declines in later layers. Together, these findings clarify how the linear accessibility of concreteness varies across LLM layers.