Training Away the Ability to Tell Evals from Deployment — LessWrong
lesswrong.com · saved by 1 readers
TL;DR: We introduce Adversarial Blinding, a training method that degrades the model's ability to distinguish between eval and deployment, with the ai…