Training
Reinforcement learning from verifiable rewards
Also called GRPO, RLVR.
What is reinforcement learning from verifiable rewards?
A training approach where the model generates several candidate answers per prompt and is rewarded according to a programmatic check of their correctness. It suits tasks where an answer can be verified automatically but the ideal reasoning that produces it cannot be written down at scale.
The classic cases are mathematics and code. Checking whether an answer equals the expected value, or whether a function passes its tests, is a few lines of code. Writing tens of thousands of ideal derivations by hand is a research project.
The reward function is the design, and it is where every failure originates. Reward only the final answer and you get a model that guesses more confidently. Reward length implicitly and you get verbosity. Whatever is measured is what improves, precisely and without regard for what was intended.
It is expensive relative to supervised training because each step involves generating and scoring multiple completions rather than one. Prove the reward function on a small model before committing a large one to it.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.