Multimodal Model Diffing for Feature Discovery and Control
Opening the black box of AI vision models to see and control what they learn
Researchers created a method to identify which internal features of AI vision-and-language models drive specific behaviors—and then selectively turn those features on or off. By comparing how a base language model changes when trained on images, they discovered sparse, interpretable features that control tasks like reading text in images and understanding spatial layouts, and showed they could degrade unwanted behaviors by up to 24% without breaking general performance.
As AI systems become more powerful and are deployed in real applications, being able to audit what's actually driving their decisions matters for safety and trust. This work provides a concrete toolkit for developers to find problematic features (like those underlying adversarial attacks) and remove them, rather than guessing blindly at how to make these systems behave. It's the difference between adjusting dials you can see versus hoping a sledgehammer approach doesn't break something important.