RLVR (Reinforcement Learning with Verifiable Rewards)

RLVR (Reinforcement Learning with Verifiable Rewards) is a post-training technique for language models introduced in the Tulu 3 paper (Lambert et al., November 2024) in which the reward signal is computed by a deterministic, programmatic verifier rather than by a learned reward model or human preference judge. The defining feature is verifiability: the reward function returns a binary or graded score that can be reproduced exactly by re-running the verifier on the model's output, with no learned model in the loop and no subjective labeling required. Typical verifiers include exact-match string comparison against a gold answer, unit-test execution for code, mathematical equivalence checking, and rule-based format validators.

RLVR is positioned as a complement to — and partial replacement for — RLHF (Reinforcement Learning from Human Feedback) and RLAIF (Reinforcement Learning with AI Feedback) in domains where ground-truth verification is cheap. The technique has been adopted rapidly across reasoning (math, code, scientific QA), tool use, and agentic settings, where the verifier can be a sandboxed test suite, a calculator, or a domain-specific oracle. It also reduces the risk of reward hacking against a learned reward model, since a deterministic verifier cannot be gradient-attacked. Limitations include the scarcity of cheaply-verifiable tasks (most real-world alignment objectives are not automatically checkable) and the gap between verifiable training tasks and the open-ended capabilities models need in deployment. The technique is also called RLVF in some early literature; the RLVR name has consolidated since 2025.