Reinforcement learning is a subfield of machine learning, where input is the features of the current state and output is an optimal action for current timestamp.

This action should maximize the expected average reward.

• Decision making is sequential
• The goal is long term

In LLMs, RL is used for alignment / fine-tuning via policy-gradient methods like Proximal Policy Optimization (PPO), GRPO, GSPO, DAPO, and preference methods like KTO.