Superposition of Safety-Critical Computations in LLM MLP Layers: A Grassmannian Subspace Analysis

We demonstrate that MLP layers 21-29 in transformer LLMs causally construct a multi-dimensional refusal subspace (AUC=0.975, causal ablation delta=-0.40), and that refusal and hallucination subspaces are encoded in distinct singular vectors of the same MLP weight matrices (Grassmann distance=0.94). This pattern replicates across Llama 3.1 8B and Mistral 7B-Instruct, suggesting a universal architectural property of transformer decoders. Research conducted with assistance from Claude (Anthropic).