Reinforcement learning used to be hard. Like, hard hard. Mostly because it was incredibly hard to evaluate agent actions, decide what to reward and punish, and how to to assign credit to which parts were involved. Now it's much easier, since the LLM can run the evaluation and determine the reward and credit assignment. We used to need humans. Now we just need agents that get it right some of the time (and a whole lot of tokens). Anyways, here's a cool article. https://www.microsoft.com/en-us/research/blog/agent-lightning-adding-reinforcement-learning-to-ai-agents-without-code-rewrites/(Shaw)