Can someone dumb to me, a generalist engineer who has a very surface level knowledge of how training LLMs work: what people were doing before and what GRPO is doing different?
They were using techniques like PPO that have a model (like a critic!) that evaluates whether the new model gives accurate responses. With GRPO, the don't have that and instead evaluate the answers based on predefined rules like coherence/ formatting. For example, for math problems, these rules will check if the answer adheres to math principles or logic!
