Can an AI learn to be honest about its own mistakes using a frozen dataset? This paper reveals a surprising 'introspective coupling' where models track their own behavior shifts even when their training labels are outdated.
http://arxiv.org/abs/2606.32038v1