Reasoned alignment: how deliberation makes language models safer
OpenAI's deliberative alignment is a new safety technique built into the o1 reasoning models. Instead of training models to memorize refusal patterns, it teaches them explicit written safety specifica…