An autoencoder learns to reconstruct its input. Sparsity limits how many features participate in each reconstruction. Applied to a language model’s internal activations, this can help separate patterns mixed together across neurons.
Anthropic’s 2023 research found features associated with patterns such as HTTP requests and DNA sequences in a small transformer. Gao and colleagues describe the tradeoff between reconstructing activations accurately and keeping the representation sparse, along with tests of feature quality.
How to read a result: a feature label is an interpretation to investigate. Check where the feature activates, where it fails to match that interpretation, and whether interventions support a claimed effect on behavior. Reconstruction and feature quality also need evaluation; one decomposition should not be treated as a complete inventory of what the model represents. Even an observed effect on behavior falls short of a complete causal explanation.
This is a research method for studying internal representations. It does not by itself make a deployed model smaller or prove that the model’s generated explanation is faithful. Explainability covers the broader question of understanding an output.
Sources
- Gao et al.: Scaling and evaluating sparse autoencoders — Describes sparse reconstruction of language-model activations and evaluation of feature quality; checked October 26, 2026.
- Anthropic: Decomposing language models into understandable components — Reports dictionary-learning features that are more interpretable than individual neurons in a small transformer; checked October 26, 2026.