Polar probe linearly decodes semantic structures from LLMs

Proposes a neural code hypothesis where semantic relations between entities are encoded by distance and direction in LLM embeddings, validated across arithmetic, visual scenes, family trees, metro maps, and social interactions.