A novel technique called "Conceptor-Based Activation Engineering".
We are bravely attempting to steer the behavior of a GPT-2XL model using Conceptors.
Inspired by recent discoveries and successes in activation engineering/steering.
by Joris Postmus & Steven Abreu (supervisor)
Conceptor Ilustration:
Paper hereExample of Activation Engineering/Steering using Activation Addition (ActAdd):
Blog post here