OpenAI releases research with Apollo AI Evals on reward-seeking behavior and Contrastive SDF method for measuring model alignment.

OpenAI details deployment simulation research using de-identified user requests to study model responses before release.