sparse autoencoder features

Sparse autoencoder features are an interpretability tool that pries apart a neural network's activations into a large set of mostly-off, individually meaningful directions. The motivating problem is superposition: a layer of a few thousand neurons seems to pack tens of thousands of concepts by letting features share neurons, so a single neuron is polysemantic and lights up for unrelated ideas. The hope is that the true features are sparse, only a handful active at once, and can be recovered.

The method trains a wide autoencoder on the model's activation vectors with a heavy sparsity penalty, learning an overcomplete dictionary of directions where each input is reconstructed as a sparse nonnegative combination of dictionary atoms. Each dictionary atom is a candidate feature; the trained encoder reports which features fire on a given input, and inspection often reveals strikingly monosemantic units, one for a specific person, a syntactic pattern, a hazard concept, far cleaner than raw neurons.

These features matter because they give a vocabulary for reading and editing internals: you can find a deception or sycophancy feature, watch when it activates, and clamp it to steer behavior. The caveats are real: feature splitting means the same concept can fragment across atoms as the dictionary grows, reconstruction is lossy, and a feature being human-interpretable does not prove it is what the model causally uses, so causal tests are needed.