Activation Engineering Research

OK, we decided to implement and then think of ideas for applying recent research to gemini! Specifically, we read these 3 papers on representation engineering and activation patches-- steering model behavior by changing their internal representations https://www.nature.com/articles/s44387-025-00031-9 https://arxiv.org/abs/2310.01405 https://arxiv.org/abs/2312.06681 One of us (sevi) did ml work to get this working in a way noone in research has yet, a sliding adjustable steering vector