Anthropic Circuits Updates January 2025: sparse autoencoder and crosscoder training optimizations

Anthropic's interpretability team published the latest setup improvements for training sparse autoencoders and crosscoders, covering the JumpReLU activation function implementation, gradient backpropagation via the straight-through estimator, and methods to reduce dead features. The setup introduces a tanh penalty term in the loss function to encourage sparsity, and adds a slight penalty for inactive features (pre-act loss); it also sets the bias initialization so that each feature is active about 10000/m of the time, and normalizes decoder weight columns to L2 norm 1 to simplify analysis code. The team said they have not yet done full ablation experiments on all design choices, and only offer this as a starting reference for external researchers training sparse autoencoders; some ideas are still in development, with more results expected in the coming months.