“Process reward models, which grade each step in a model's chain-of-thought, are a key component for training models with advanced reasoning, as demonstrated in a preliminary paper by OpenAI.”