RLVR: training AI with rewards you can verify
RLVR (reinforcement learning with verifiable rewards) is a training method in which a model is rewarded only when its answer passes an automatic check: an exact maths result, code that passes its tests, an instruction that was followed. It has played a central role in the rise of reasoning models.
Data as of October 22, 2026
Key figures
| Level | Advanced |
|---|---|
| Origin | 2024 |
| Introduced by | Lambert et al. (Tülu 3) |
The principle
The idea is to swap the learned “judge” for an automatic grader: right answer, reward; wrong answer, nothing.
In RLHF, a reward model estimates what people would prefer. That learned judge can be fooled: the model being trained sometimes learns to please it without actually answering well. RLVR sidesteps the problem in domains where correctness can be checked. The model produces several attempts, a verification function gives each one a reward, often simply pass or fail, and a reinforcement learning algorithm (PPO, GRPO and others) makes the successful attempts more likely.
A concrete example
Problem: “A train leaves at 2:10 pm and arrives at 5:45 pm. How long is the journey?” The model writes out its reasoning, then gives a final answer in a required format. A script compares that answer with the expected solution, 3 hours 35 minutes. Whether the reasoning was long or short does not matter: only the correctness of the result counts. In programming, the check means running unit tests on the generated code.
Origins
The name was introduced in November 2024 in “Tülu 3: Pushing Frontiers in Open Language Model Post-Training” by Nathan Lambert and colleagues at the Allen Institute for AI, who present it as a new method applied after supervised fine-tuning and DPO. In January 2025, DeepSeek described training DeepSeek-R1 with rule-based rewards (answer accuracy and format), deliberately avoiding neural reward models, which it considered vulnerable to reward hacking in large-scale reinforcement learning.
Why it matters, and common misconceptions
RLVR is one of the engines behind reasoning models: rewarding only the correct result pushes the model to develop long chains of thought, double-check itself and fix its mistakes.
- “A verifiable reward can’t be gamed.” If the tests are incomplete, the model can satisfy them without solving the problem, for example by hard-coding the expected outputs.
- “It works for everything.” Fields with no objective check, such as style, advice or creative work, are a poor fit.
- “The visible reasoning must be right.” Only the final result is checked; clumsy reasoning can still land on the correct answer.
Related terms
- Reasoning model · A language model trained to think step by step before answering: slower and pricier, but stronger at maths, code and logic.
- Reward model · An auxiliary model that scores an answer by predicting human preference, used as the training signal for another model.
- RLHF (reinforcement learning from human feedback) · A training method where people compare model answers, and the model learns to favour the responses humans judge helpful and safe.
- Test-time compute · The computing power a model spends while answering, for instance to think for longer or to compare several candidate solutions.
- Chain of thought · Getting a model to write out intermediate reasoning steps before its final answer, which helps on multi-step problems.
Full glossary
Frequently asked questions
What is the difference between RLVR and RLHF?
In RLHF the reward comes from a model trained to imitate human preferences, in other words a judgement. In RLVR it comes from a program that checks the answer objectively. RLVR is more reliable wherever such a check exists, but it does not apply to everything.
Can RLVR be used to teach email writing?
Only partly. You can check precise constraints such as length, a required phrase or a format, but not the overall quality of the tone or argument. For that you fall back on human judgements or a reward model.
Does RLVR make models smarter?
It makes them much more reliable on verifiable tasks. Researchers do debate, however, how much it creates genuinely new abilities versus surfacing abilities the base model already had more consistently.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.