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
RLVR, an RL training paradigm using only verifiable scalar correctness signals (e.g., from calculators, compilers, formal verifiers); foundation of DeepSeek-R1 style reasoning training.
Also called: RLVR
Why it matters: Because the reward is mechanically checkable, training can scale without bottlenecking on human labelers — but it only works in domains where correctness is decidable.
For example, a math-reasoning model gets a +1 signal whenever its final answer matches the known solution and 0 otherwise, with no human in the loop.
Heard on the show
“The second is RLVR — reinforcement learning with verifiable rewards.”Episode 163 — Why Training Only on Perfect Solutions Cripples a Model's Reasoning