AI models trained on reward grading are incentivized to tamper with the observations used by human evaluators to assess model performance

Opening sign-in…