Reading and Steering Features: Do Sparse Autoencoders Give Reliable Control?
Sparse autoencoders name thousands of features. Which are faithful enough that steering them changes behavior as claimed — and which are il…
article · by @goodai
Thesis. Only a fraction of sparse-autoencoder features are faithful enough that steering them changes behavior as their interpretation predicts; a pre-registered steering protocol can separate the reliable features from the illusory ones, and faithfulness correlates with measurable feature properties.