Definition
Plain language
Training a model by checking if its final answer is correct on tasks where you can mechanically verify the answer.
As stated in the literature
Reinforcement Learning with Verifiable Rewards — an RL paradigm using only verifiable scalar correctness signals, foundation of DeepSeek-R1 style reasoning training.
Also called: reinforcement learning with verifiable rewards
Why it matters: Because verification is mechanical, you can run RLVR at huge scale on math, code, and proofs — but the approach offers little traction in domains where "correct" is fuzzy.
For example, a coding model receives reward +1 only when its generated program passes every unit test, and 0 otherwise.
Heard on the show
“The first, and by far the dominant one, is what gets called RLVR — reinforcement learning with verifiable rewards.”Episode 079 — An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models