강화학습 지식 라이브러리
How feedback changes AI behavior: foundations, policy optimization, human preferences, verifiable rewards and agent training. Follow the sequence and test each claim.
Detailed technical articles are available in English and Chinese. This locale shows the English article with localized navigation.
Local demos calculate toy objectives and state transitions. They do not train an LLM. Real-model protocols require a separate compute environment and may have cost or data-transfer implications.
- 1
- 2
Exploration, credit assignment & reward design
Make useful behavior discoverable and measurable.
Study this chapter → - 3
Policy gradients, baselines & actor–critic
Understand how rewards change action probabilities.
Study this chapter → - 4
PPO, clipping & on-policy data
Limit an update incentive without assuming a hard trust region.
Study this chapter → - 5
SFT → RLHF → DPO: different training signals
Choose the signal that matches the available feedback.
Study this chapter → - 6
GRPO, verifiable rewards & reasoning
See how an outcome signal can reinforce successful solution behavior.
Study this chapter → - 7
Agent RL: train a policy inside a harness
Make the environment, tools and success criterion reproducible.
Study this chapter → - 8
Does RL actually improve capability?
Make a falsifiable claim with independent evidence.
Study this chapter →