Glossary Entry

Reinforcement Learning from Verifiable Rewards

Post-training a language model with reinforcement learning against rewards that a program can check automatically, such as a correct final answer or passing tests, instead of a learned human-preference reward.

RL Training LLMs

Also called: RLVR, verifiable rewards

Reinforcement learning from verifiable rewards (RLVR) keeps the machinery of RLHF, a policy optimised with an algorithm such as PPO or GRPO, but swaps the reward source. Instead of a reward model fitted to human comparisons, the reward is computed by a checker: does the final answer match the reference, does the code pass its tests, does the output follow the required format. That makes the reward cheap, noise-free, and hard to game on tasks that have checkable solutions.

It is the recipe behind the reasoning models, where a model trained this way on mathematics and code learns to produce long chains of thought without being shown any. Its limit is the checker: tasks without an automatic verifier, such as judgement calls or open-ended writing, need a different reward, which is why preference rewards and, more recently, calibration-based rewards continue to exist alongside it.