The State of Mechanistic Interpretability
for AI Safety

Methods · Sparse Autoencoders / SAE Variants and Architectural Improvements

Batch Top-k SAEs

Survey sections: §6.3 SAE Variants and Architectural Improvements

Batch Top-k SAEs relax the fixed per-example sparsity constraint of standard Top-k SAEs by enforcing the sparsity budget across an entire batch instead (Bussmann et al., 2024).

Assumptions and requirements

This loose constraint allows the model to allocate representational capacity adaptively across samples.

Intended applications

More recent Batch Top-k SAEs allocate sparsity adaptively across samples, which allows more active features on complex inputs and fewer on easy ones (Bussmann et al., 2024).
During inference, Batch Top-k SAEs instead use the JumpReLU function with an inferred global threshold to determine which features are active for each input, so that an input’s activations do not depend on the rest of the batch.

Related methods

Compare side by side →

Used in safety applications

Run it

Key papers

Cited in §6.3.

Also discussed in: §6.5.1 Dead Features · §9.1 Dictionary Learning Methods

Other methods in Sparse Autoencoders / SAE Variants and Architectural Improvements

Sparse Autoencoders (SAEs) · Gated SAEs · JumpReLU SAEs · Top-k SAEs