Is Masking a Way to Stimulate Pre-Trained Large Language Models?

Masks could be used to probe pre-trained models, perhaps for specific parts of the model that are responsible for interrogating certain characteristics of the input data. That certainly connects, I think, to other work that has investigated ways of steering models or intervening on the concepts and things that they look at. So this is really cool because it's kind of like this idea that the pre-trained model, you know, maybe has a store of many different features, and by selecting different combinations of these features, you can adapt them to and specialize them to a different path.

Transcript

Play full episode

Transcript

Episode notes

The AI-powered Podcast Player

Save insights by tapping your headphones, chat with episodes, discover the best highlights - and more!

Get the app