Skip to content
BENIBIX Learn

RL with Verifiable Rewards (RLVR)

AI💬 LLMsLanguage Model Training

Also known as RLVR, Reinforcement Learning with Verifiable Rewards, Verifiable Rewards

Points for a model each time a check shows it got the right answer.

Think of it like

It is like a quiz with an answer key. A sticker comes only when your answer matches the key.

Example

A model works through 1,000 math problems. A short script checks each final number against the answer key, and only the right ones earn a reward.

Why it matters

A check by a program is cheap and hard to argue with. That makes long training on math and code possible.

Definition

Reinforcement learning with verifiable rewards, or RLVR, trains a model on tasks a program can check. The model tries a problem, such as a sum or a coding task. A checker compares the final answer with the key, or runs the code's tests. It gives a reward only for a pass, so no person or reward model has to judge. The name comes from a late 2024 paper. The method is used to train reasoning models on math and code.
The model answers and a program checks it. Only a right answer gets a reward.

Common mistake

Using it where there is no single right answer. A script cannot mark a poem or an essay.

Key takeaways

  1. 1
    A checkable answer. Math, code and puzzles.
  2. 2
    A program judges. No human rater.
  3. 3
    Pass or fail. Reward only for a match.

Compare

Learn it properly