SAEVerbalizer Explains Sparse-Autoencoder Features From Decoder Directions Instead of Observed Behavior
Current SAE feature explanation depends on external observation — collecting behavioral evidence at scale, which is both expensive and yields superficial descriptions. SAEVerbalizer injects SAE decoder directions directly into an LLM's representations and fine-tunes the downstream layers to emit natural-language explanations of the injected feature, so explanation afterward needs only the decoder direction. The learned capability generalizes to unseen features, transfers across separately trained SAE dictionaries and, with a lightweight adapter, to SAE features from different LLMs; intervention experiments show injecting multiple directions produces a combined-meaning explanation and reversing a direction shifts the meaning correspondingly.
↳ Follow the thread