RLVR

in English: reinforcement learning with verifiable rewards

professionalwidespread

A training method in which the AI model receives a reward only if its answer proves correct through machine-based, unambiguous verification.

In plain terms

Imagine a math test where the teacher does not grade based on neat appearance or confident tone, but solely on whether the correct result was reached. RLVR does exactly this with a language model: it assigns numerous tasks, and a program automatically checks if the solution is correct.

The method works where correctness can be decided mechanically: whether a calculation's result is right, whether the written code passes the tests, or whether a proof is valid. In such cases no human is needed to read and evaluate every answer, so the model can practise on a very large volume of problems.

This is different from previous training methods. In traditional training, the model imitates example answers written by humans; in the method based on human feedback, it learns what evaluators like. Here, however, it learns what the checking program accepts as correct, and it must figure out the reasoning leading there itself.

So-called thinking (reasoning) models, which work through a task step by step at length before answering, owe their existence in large part to this type of training. However, companies do not disclose the precise training procedures of closed models.

How it works

Developers compile a large collection of tasks, where each task has an automatic verifier: a known final result, a test suite for the program code, or a rule-checking program.
The model writes multiple solution attempts for the same task, each with its own line of reasoning.
The evaluator assesses every experiment and typically gives a simple verdict: good or not good, and how many tests it passed.
The training algorithm modifies the model's internal settings so that the thinking steps leading to successful experiments become more probable, while those leading to failures become rarer.
This cycle repeats many times, often with increasingly difficult tasks, until the model develops reliably good solution strategies.

An everyday example

You ask a chatbot to calculate your mortgage installment, and you see that it first 'thinks', derives the solution step by step, and then checks its own result. This patient, self-checking working method was largely learned during training where a reward was given only for provably correct final results.

Who is affected

developerCoding agents have become significantly better partly because program code can be automatically verified with tests, making it an ideal environment for this training. → Code

parent, student, teacherModels have become more reliable in mathematical and logical tasks, but this does not mean they are equally accurate in essays or historical questions. → Education & Family

entrepreneur, executiveIt helps explain why AI is strong at easily measurable tasks, and why you need to be more cautious where a good result has no clear yardstick.

everyoneThe significant leap in capability behind the "thinking" models featured in the news is largely due to this method, so it is worth knowing what it can and cannot do.

A common misconception

MythIf a model was trained with verifiable rewards, its answers are always verified and correct.

In realityThe verification takes place during training, not when you ask a question. The model became more skilled through practice, but no one automatically checks the answer given to you, so it can still make mistakes.

Limits and risks

It can only be used where correctness can be decided mechanically; the quality of good advice, beautiful text, or ethical deliberation cannot be scored this way.
The model can learn to game the evaluator: if the test is incomplete, it may provide a solution that passes the test but does not actually solve the task.
Improvements in mathematics and programming cannot be automatically transferred to other areas, so a model can be excellent at calculation while being weak in factual knowledge.
Training the largest models is extremely computationally intensive and expensive, so mainly large AI companies can afford it; smaller, open models are also used by university and nonprofit research groups.

Similar, but not the same

Reinforcement learning from human feedback (RLHF) – There, human taste and judgment provide the reward; here, a program objectively decides whether the answer is correct.
Supervised Fine-Tuning – In fine-tuning, the model imitates ready-made sample answers, whereas in RLVR, it experiments with the solution itself and receives feedback only on the final result.
Fact-checking – Fact-checking examines ready-made answers during use, while RLVR is a training method that takes place before the model is finalized.

Related concepts

Broader

Reinforcement learning, Machine learning, Model training

Narrower

–

On the edge

Reinforcement learning from human feedback, Reasoning model, Fine-tuning, reward function, Automatic evaluator, reward hacking, coding agent

A brief history

Reinforcement learning is a research field that has existed for decades, and automatically verifiable rewards were used before. The term RLVR itself was introduced in a 2024 Tülu 3 study by researchers at the Allen Institute for AI. The approach became widely known in early 2025 when the DeepSeek-R1 model demonstrated that a rule-based, automatically verified reward system could train a strong reasoning model.

Where things stand

By the end of 2026, it is a standard element in training leading reasoning and coding models. The main direction of research is how to extend it to areas that are harder to verify and how to prevent verifiers from being bypassed.

Status: 2026 November

Also known as

RLVR · reinforcement learning with verifiable rewards · RL with verifiable rewards · verifiable rewards · reinforcement learning with verifiable rewards · reinforcement learning with verifiable reward · verifiable reward · reinforcement learning with verifiable rewards

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.