Cross-Domain Activation Analysis

Gemma-2B SAE | 8 domains × 4 layers | Per-question PCA-10 analysis
Phase 1: Domain Similarity
Similarity Matrix
Ward Clustering
Phase 2: SVM Transfer Classification

Method

Train LinearSVC on PCA-10 to separate spatial vs language. Apply classifier to other 6 domains. Result = % classified as "spatial-like".

Method Comparison (Layer 22)
Key Findings