A neural network’s internal activation at a given layer is a single dense vector, and any one direction in that vector space often turns out to encode several unrelated concepts at once (a phenomenon usually called superposition) rather than one direction per human-interpretable idea. A sparse autoencoder (SAE) addresses this for mechanistic interpretability by learning to reconstruct that dense activation from a much larger, sparse intermediate representation — many more “features” than the original activation had dimensions, but only a small number of them active for any given input. Concretely, an activation is encoded into sparse feature strengths and reconstructed as
where each (a column of the decoder matrix ) is that feature’s own direction in activation space.
Because each SAE feature tends to activate for a narrower, more specific pattern than a raw activation direction does, individual features are more often directly interpretable, and specific features (or coherent groups of them, sometimes called circuits when they co-activate and connect across layers) can be identified and manipulated for a target behavior with less interference from unrelated concepts than intervening on a raw activation direction would cause.
Related papers
- Saadatinia, Razmara, Aryashad, Abbasi, and Azizi, “CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits”, arXiv:2608.05732 (2026) — uses SAE features co-activating across layers to isolate multi-layer circuits for activation steering.