We just posted a blog + paper on a a simple but effective approach to model honesty called "Confessions"
TL; DR: normal RL training rewards for high performance on a task. Confession training is a separate phase that rewards only for honesty. Test look promising!
More:
TL; DR: normal RL training rewards for high performance on a task. Confession training is a separate phase that rewards only for honesty. Test look promising!
More:
In a new proof-of-concept study, we’ve trained a GPT-5 Thinking variant to admit whether the model followed instructions.View quoted post on X
This “confessions” method surfaces hidden failures—guessing, shortcuts, rule-breaking—even when the final answer looks correct.
openai.com/index/how-conf…
