This research systematically analyzes Sparse Autoencoders (SAEs) as tools for extracting human-interpretable representations from Large Language Models, essential for applications requiring transparency and control.

• Evaluates optimal model-layer selection and scaling properties for feature extraction
• Investigates different SAE architectural configurations including width and pooling strategies
• Demonstrates SAEs can extract interpretable features useful for classification tasks
• Explores transferability of features across models and tasks

Medical Significance: In healthcare applications, this approach enhances explainability and transparency of AI systems—critical for clinical decision support, medical diagnostics, and ensuring regulatory compliance where interpretability is non-negotiable.

Sparse Autoencoder Features for Classifications and Transferability