AI glossary · Advanced · Reinforcement learning · Language models
RLVR
RLVR (Reinforcement Learning with Verifiable Rewards) is a method of training AI models with reinforcement learning, using as the reward criteria that can automatically check right and wrong, such as matching against the correct answer or running tests.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Perception & action
- Added
- 2026-10-08
How it works
The model is given problems whose answers can be checked, such as math problems or coding tasks, and generates several answers itself (rollouts). A verifier then compares the final answer with the correct one or runs the code against tests, giving a high reward if it is right and a low reward if it is wrong. The model is updated so that the approaches that produced high-reward answers appear more often. Repeating this process leads the model to find better ways of solving problems on its own.
Why it matters
Reward models trained on human preferences (the RLHF approach) are expensive to build, and models can easily exploit their weaknesses. Because RLVR scores precisely by rules, these problems are reduced, and it is easy to run automated training on large numbers of problems. It is especially effective on problems that require multi-step thinking, such as math and coding, so it is regarded as a key technique for building reasoning models that think through a long CoT before answering.
Things to know
It can be applied directly only in areas where answers can be checked automatically. It is hard to apply to tasks such as writing or counseling, where the criteria for a good answer are vague. Also, if the verifier has loopholes, for example if the tests are weak, reward hacking can occur, where the model passes the tests without properly solving the problem. When only the final answer is scored, there is no separate guarantee that the intermediate reasoning is correct.
Example
As in the paper 'TRACE: Rollout-based QAT (quantization-aware training) for FP4 reinforcement learning of MoE language models', covered on this site, the computational cost of rollouts (the stage where the model generates answers itself) is a heavy burden when training language models with reinforcement learning, and research to reduce it is active. Because RLVR is built around generating large numbers of such rollouts and scoring them automatically, the cost issue is especially important. In real services, it is used to train reasoning models that solve math problems step by step and coding agents that write code that passes tests.