LLM Representations Show Transient Human-Like Perceptual Geometry

A new study on arXiv (2605.27970) reveals that large language models (LLMs) trained solely on text develop internal geometric structures resembling human perceptual organization across domains like color, pitch, emotion, and taste. Researchers analyzed layer-wise emergence in residual streams of open-weight transformer architectures, finding three key results: geometric structure emerges across multiple perceptual domains without direct perceptual supervision; these domains have distinct emergence profiles with domain- and model-specific alignment to human baselines; and the structure is transient, appearing only in certain layers. The work builds on prior evidence of rich geometric structure in LLM embeddings, suggesting that textual training alone can encode aspects of human perception.