RLVR
Reinforcement Learning with Verifiable Rewards is a training methodology that optimizes models by automatically checking logical outputs against absolute tru...
Think of It Like This
Like teaching a student math by having a computer instantly verify if their final answer matches the known solution, rather than asking a human grader.
Unlike RLHF which relies on subjective human preferences, RLVR is used in domains with objective ground truth, such as writing executable code, passing unit tests, or proving mathematical theorems. Because the reward signal is perfectly reliable and incredibly cheap to generate, models can undergo massive, automated self-improvement cycles.