Everything tagged model evaluation.
The method replays about 1.3 million de-identified ChatGPT conversations against a candidate model, and caught a reward-hacking behavior OpenAI calls…