Reinforcement Learning with Verifiable Rewards (RLVR)

RLVR is a training methodology for language models in which reinforcement learning rewards are derived from objective, programmatically verifiable criteria rather than from a neural network trained on human preference annotations. The term was widely adopted following the release of DeepSeek-R1 in January 2025, though the underlying principle appeared in earlier work on code generation and mathematical reasoning.

In practice, RLVR applies policy-gradient algorithms — most commonly Group Relative Policy Optimization (GRPO) or a variant of PPO — where the reward function is a deterministic checker: for a math problem, the model's final numerical answer is compared to a ground-truth value (correct = +1, incorrect = 0); for code, the generated program is executed against hidden unit tests and scored on pass rate. This structure directly avoids reward hacking — the tendency of learned reward models to be exploited by surface-level linguistic patterns that score well without being genuinely correct.

RLVR has become the dominant approach for training reasoning-specialized models because verifiable domains provide abundant, reliable training signal without per-instance annotation cost. The DeepSeek-R1 technical report showed that a base model trained with RLVR on mathematical and coding problems spontaneously developed extended chain-of-thought reasoning — including self-correction and multi-step exploration — without any supervised reasoning traces. Similar results were replicated by Qwen, Kimi, and several academic groups within months of that release.

As of mid-2026, RLVR is a core training stage for frontier reasoning models from most major labs. Research is expanding the verifiable reward paradigm beyond math and code to formal theorem proving with Lean 4 proof checkers, structured scientific data generation, and database query synthesis. Open-source RLVR training frameworks such as OpenRLHF and verl have lowered the barrier to replication for smaller research teams.

Example

“A reasoning model trained with RLVR on competition mathematics datasets is rewarded only when it produces the exact correct numerical answer, causing it to learn longer, self-checking reasoning chains rather than surface-level pattern-matching shortcuts.