Model can say it's editing weights for reason X, and do so while simultaneously modifying weights for silent reason Y. Its edit will pass the evaluation, but it has silently injected new behavior Y in ways that might be very difficult to detect.
I understand, it has underlying alternative motives and knows to be deceptive based on all of the material is has been trained on. It's behavior Y is only detected when its too late.
I understand, it has underlying alternative motives and knows to be deceptive based on all of the material is has been trained on. It's behavior Y is only detected when its too late.