SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization
SAEVerbalizer turns sparse-autoencoder decoder directions into natural-language feature explanations. It may reduce behavior-sampling work when inspecting model representations, but evidence is experimental.
SAEVerbalizer injects **SAE decoder directions** into a language model and fine-tunes downstream layers to explain the represented features. The learned verbalizer generalized to unseen features and separately trained SAE dictionaries.
For model inspection tooling, this suggests a direct path from internal feature vectors to readable descriptions without collecting behavioral examples for every feature. A **lightweight adapter** also transferred the method to SAE features from other language models.
SAEVerbalizer injects **SAE decoder directions** into a language model and fine-tunes downstream layers to explain the represented features. The learned verbalizer generalized to unseen features and separately trained SAE dictionaries. For model inspection tooling, this suggests a direct path from internal feature vectors to readable descriptions without collecting behavioral examples for every feature. A **lightweight adapter** also transferred the method to SAE features from other language models. The paper reports semantic shifts when directions are combined or reversed, but the supplied material gives no accuracy figures or external validation. Natural-language explanations can therefore aid investigation without proving that a feature has been interpreted correctly.
This adds a potentially scalable inspection layer that verbalizes internal SAE directions without requiring feature-by-feature behavioral datasets, including transfer across dictionaries and models. It makes broad feature triage more practical, but the reported semantic changes under direction composition and reversal reinforce that readable explanations are hypotheses for validation, not proof that a feature’s function has been identified.