Did grading make the model more honest, or just less obedient? Would the graders also ignore a harmless instruction like “reply in Spanish”?
Did grading make the model more honest, or just less obedient? Would the graders also ignore a harmless instruction like “reply in Spanish”?