#feature learning

Sparse Autoencoders Show Feature Starvation and Structural Instability

Two papers reveal fundamental limits in SAE interpretability methods for large language models.