SALVE: Sparse Autoencoder-Latent Vector Editing for Mechanistic Control of Neural Networks

We can now "see" what neural networks are doing internally: SALVE uses sparse autoencoders to identify specific features that networks rely on for decisions, making it possible to understand which patterns a model has learned—a major step toward opening the "black box" of AI systems.

Discovered features can be validated and visualized: The paper introduces Grad-FAM, a method that grounds these internal features in actual input data, allowing researchers to confirm that identified features correspond to meaningful patterns rather than statistical artifacts.

AI systems can be precisely edited without retraining: By directly modifying network weights based on identified features, SALVE enables permanent behavioral changes—such as removing unwanted biases or capabilities—without the time and computational cost of retraining entire models.

The framework reveals how dependent models are on specific features: The critical suppression threshold (α_crit) quantifies how much a model relies on particular features for each class, providing diagnostic information about potential vulnerabilities or robustness issues in deployed systems.

The approach works across different AI architectures: Validation on both convolutional networks (ResNet-18) and transformer models (ViT-B/16) suggests this interpretability and control framework could be broadly applicable, supporting more responsible deployment of diverse AI systems.