Fetching the paper…

Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization · Around