Improving Model Safety Behavior with Rule-Based Rewards

OpenAI outlined Rule-Based Rewards, a training method that uses explicit safety rules instead of relying mainly on human feedback.

The company said the approach has been part of its safety stack since GPT-4 and GPT-4o mini.

It could make safety training faster, easier to update, and more scalable than traditional RLHF.

OpenAI says the method may also help models refuse harmful requests more consistently.