Seonglae ChoFeature Steering
SAE-based vectors enable more fine-grained control. Because SAE vectors are designed to be sparse, they minimize the impact on other behaviors, allowing for more precise adjustments.
Contrastive Activation
Neuron SAE 와 달리 두 가지 시나리오 모델 활성화 차이를 직접적으로 겨냥해서 반영한다. 이진적이거나 큰 행동 변화를 포착하는 데 유용하지만, 그만큼 정밀한 조절은 어렵고 여러 연관된 행동을 함께 조절할 가능성이 있다. 즉 다른 행동에 영향을 미칠 가능성도 있다.
Steering Vector Usages
Feature Steering Notion
![[Full Post] Progress Update #1 from the GDM Mech Interp Team — LessWrong](https://www.notion.so/image/https%3A%2F%2Fres.cloudinary.com%2Flesswrong-2-0%2Fimage%2Fupload%2Fv1497915096%2Ffavicon_lncumn.ico?table=block&id=bcba6361-8c67-4b1b-ad8d-f1d0466f8903&cache=v2)
![[Full Post] Progress Update #1 from the GDM Mech Interp Team — LessWrong](https://www.notion.so/image/https%3A%2F%2Fres.cloudinary.com%2Flesswrong-2-0%2Fimage%2Fupload%2Ff_auto%2Cq_auto%2Fv1%2FmirroredImages%2FC5KAZQib3bzzpeyrg%2Furkzhkxl8td0ovllz8zz?table=block&id=bcba6361-8c67-4b1b-ad8d-f1d0466f8903&cache=v2)