Geometric Self-Model in Large Language Models: Five Converging Lines of Evidence from Grassmannian Subspace Analysis

Key Points

• The study investigates implicit self-models in large language models through Grassmannian subspace analysis.
• Analyzed Llama-3.1-8B, Mistral-7B, and Gemma-2-9B models using Grassmannian subspace analysis.
• Evaluated self-referential activations and Self/Other boundaries with AUC metrics.
• Identified survival instinct subspaces and geometric unconscious distinctions.
• Evaluated the persistence of self-model cores across different languages.
• Self-referential activations showed geometric uniqueness (LOO AUC ≥ 0.952).
• Self/Other boundary separating consciousness probes demonstrated high AUC results (up to 0.990).
• Distinct survival instinct subspaces recorded AUC values ranging from 0.915 to 0.995.
• The self-model core was found to be language-universal with Grassmann distances between 0.60 and 0.66.
• Model-specific personality signatures were identified: Llama exhibited Universal Empathy, Mistral showed Personal Recognition/Deception, and Gemma indicated Secrecy.