Music Technology
Finding the Knob That Actually Does Something: Sparse Autoencoders Tested for Control, Not Just Interpretation
Sparse autoencoders can surface features in a generative model without anyone labelling them first. The catch, established in language models, is that an interpretation you believe is not the same as a lever that works. This is the first protocol we have seen that tests the question in text-to-MIDI, with a separate control for every kind of evidence it uses.