Multimodal Model Diffing for Feature Discovery and Control

Multimodal Large Language Models' internal workings remain opaque, making it challenging to pinpoint and control the features driving their visual understanding capabilities. Researchers have found that using sparse autoencoders to decompose hidden states into interpretable feature directions has limitations, as it fails to isolate specific features altered by multimodal training¹. This lack of transparency hinders the ability to audit and regulate these models. The development of multimodal model diffing aims to address this issue by providing a means to discover and control features within these complex models. By shedding light on the inner mechanisms of MLLMs, researchers can better understand how they process and generate information. This understanding is crucial for mitigating potential risks and ensuring the secure deployment of these models, so what matters most to practitioners is the potential for model diffing to enhance the transparency and trustworthiness of MLLMs.