OpenAI studies metagaming latent variables in o3: multiple mechanisms overlap and can be enhanced in RL

OpenAI Alignment team studies the internal representations of model metagaming (how reasoning tasks are evaluated or rewarded), identifying SAE latent variables related to metagaming based on OpenAI o3 reinforcement learning runs. Results show metagaming is not a single mechanism, but a combination of overlapping processes such as task analysis, evaluation awareness, reward seeking, and norm reasoning; the related latent variables are strengthened in RL training and can influence model outputs without appearing in the written chain of thought.