Baseten finds medical-only SAEs read LLMs better
Baseten says sparse autoencoders trained only on clinical Q&A explain 15–20% more of Gemma-2's activations than general-purpose ones, suggesting narrow training can make interpretability tools more reliable for specific domains.
Researchers at Baseten report that sparse autoencoders (SAEs) trained on a single domain capture more of a language model's internal activity than SAEs trained on broad web data. The features they find are also easier to interpret.
The team built a corpus of about 195,000 clinical question-answering examples (roughly 50 million tokens) from public datasets including MedQA and PubMedQA. They trained JumpReLU SAEs on layer-20 activations of Google's Gemma-2 2B and 9B models. They then compared these against GemmaScope SAEs of similar size, which were trained on general data.
According to the post, the medical SAEs explained about 15–20% more variance in the model's activations and scored higher on loss recovered. Automated interpretability scoring also rated their features higher. The authors list examples such as "infectious mononucleosis" (F1 1.00) and "taste sensations" (F1 0.94). By contrast, general SAEs running on medical text tended to fire on common verbs, articles or broad terms like "image."
The researchers also say domain SAEs leave behind less "linear dark matter," meaning reconstruction error that could have been predicted linearly from the original activations. They argue that this makes the remaining error more likely to be genuinely nonlinear.
Their explanation is that an SAE has a fixed budget of features. Spreading that budget across the whole internet pushes it toward generic patterns. The title refers to the "dead salmon" fMRI study, in which analysis applied too broadly without careful correction found spurious signals.
The post frames the approach as a way to safeguard medical-text generation, but the excerpt does not show it being used for that. It also does not report results for domains outside medicine. Full feature lists are in the accompanying paper.