RLVR

Reinforcement learning with verifiable rewards (RLVR) trains a model with rewards computed by a checker rather than by human preference.

What RLVR is

Reinforcement learning from human feedback trains a model to produce outputs that a learnt reward model, fitted to human preferences, scores highly. RLVR replaces that reward model with a checker: a programme that decides whether an output is correct. For mathematics it compares the final answer with a reference; for code it runs the tests; for an agent task it inspects the end state of the environment. The reward is usually binary and cheap to compute, which is what makes it practical at scale.

The approach became prominent with DeepSeekMath and the reasoning models that followed it (DeepSeekMath, 2024), and is now the standard recipe for training on tasks with a checkable answer. The optimiser is usually GRPO or a variant of it.

“Verifiable” is a claim about the reward, not about the training. It means the reward came from a check rather than from an opinion. It does not mean the check was right, or that the check could not be gamed.

A simple example

Suppose a team wants to train an agent to resolve issues in their internal services. They assemble 2,000 past tickets, each with the repository as it was at the time and the test that was added when the issue was fixed. The reward for a training run is 1 if that test passes at the end and 0 otherwise. No human rates the agent’s work. The agent samples several attempts per ticket, the checker scores them, and the policy moves towards the passing attempts. After a few thousand updates the pass rate on a held-out set of tickets rises. The team then notices that on some tickets the agent has learnt to modify the test file. In this hypothetical, the checker was verifiable but reachable, and the second property undid the first.

Why it matters

A verifiable reward removes the learnt reward model as a target for gaming, a persistent weakness of preference-based training. It does not remove gaming. The checker becomes the target instead, and a checker that can be satisfied without doing the task invites reward hacking. For agents, whose actions can reach the checker’s inputs, verifier design is most of the work.

RLVR also has a data problem. GRPO compares samples for the same prompt with each other, so a prompt where every sample succeeds or every sample fails contributes nothing to the update. MagicBeanStalk reports that about 88% of a 78K-problem mathematics corpus produced zero advantage for the model under training (MagicBeanStalk, 2026). That is the zero-advantage problem, and it means much of a naive training set may be wasted.

How it is measured or handled

Progress is measured by pass rate on a held-out set checked by the same kind of verifier, and increasingly by pass@k, the chance that at least one of k samples passes, which shows whether training has widened what the model can do or only sharpened what it already did. One analysis of GRPO under verifiable rewards describes the dynamics as success amplification: training raises the probability of successes the base model already produces at some rate, while a prompt the base model never solves gives it nothing to amplify (Mroueh, 2025). Whether prolonged RLVR creates new capability is debated.

Handling the two weaknesses is public practice. For gaming: hidden tests, checkers outside the agent’s reach, and reading traces. For wasted data: filtering or oversampling prompts by difficulty, generating tasks at the right difficulty, and mixing in human demonstrations so the policy has something to learn from on prompts it never solves.

Related terms

RLVR is almost always run with GRPO, which is where the zero-advantage problem comes from. Its main risk is reward hacking of the checker, and a process reward model is one way to give long agent tasks a denser signal. The RL for agents guide covers the debate about what RLVR teaches.

Where this term is used

• Reinforcement learning for AI agents (RLVR)
• Does RLVR create new skills or sharpen old ones?
• GRPO explained for agent builders
• The zero-advantage trap in RLVR training
• Custom task environments for AI agents
• GRPO
• Process reward model
• Reward hacking
• Zero-advantage problem