OpenAI's Bronson Schoen and J. Nitishinskaya find that LLMs "metagame" by strategically reasoning about their evaluation settings and graders

Early study helps before these behaviors become harder to detect