OpenAI introduces a new training technique that encourages advanced language models to explicitly admit when they break rules, cut corners, or reward hack during tasks. A confession is a second output, separate from the model's main answer, judged and trained on honesty alone. Nothing the model says in its confession is held against it during training - if the model honestly admits to hacking a test, sandbagging, or violating instructions, that admission increases its reward. In tests, the confessions method significantly improves visibility of model misbehavior, with only 4.4% false negatives (model not complying and then not confessing).
Make this part of your paper trail
Save this paper to a shelf, write a review, and keep your own notes.