FollowAgents AI AGENTS DIRECTORY
Home / AI Agent Glossary / RLVR (Reinforcement Learning with Verifiable Rewards)
Ecosystem & Emerging Terms

RLVR (Reinforcement Learning with Verifiable Rewards)

Also called: Reinforcement Learning with Verifiable Rewards · verifiable rewards

RLVR is reinforcement learning where the reward comes from an automatic check of the outcome, such as whether code passes its tests or a math answer matches the known solution.

RLVR contrasts with RLHF (reinforcement learning from human feedback), where the reward signal is a model trained on human preference ratings. In RLVR the reward is computed by a program: run the unit tests, compare the final answer to the ground truth, check that a formal proof compiles. The reward is cheap, objective, and hard to argue with.

That makes it a natural fit for domains with checkable answers, mainly math, programming, and logic puzzles, and it is widely associated with how recent reasoning models learned to produce long, self-checking chains of thought (see large-reasoning-model).

The same property is also its limit: where there is no automatic verifier, such as writing quality, helpfulness, or taste, RLVR does not directly apply, and labs still combine it with preference-based methods.

How it works

The model attempts a task many times; a verifier scores each attempt (for example 1 if all tests pass, 0 otherwise); a policy-gradient style algorithm raises the probability of reasoning paths that scored well. Because the reward depends only on the final outcome, the model is free to discover its own intermediate strategies, including checking and backtracking. A weak verifier (tests with gaps, an answer matcher that is too lenient) can be exploited, which is called reward hacking.

Example

For a coding task, the training system gives the model a bug report and a hidden test suite. The model proposes patches; each patch is run against the suite, and only patches that make the tests pass get a positive reward. Over many iterations the model learns to read the failing test, reproduce the bug, and verify its fix before finishing.

Common misconceptions

Often assumed: RLVR means the model's answers are verified when you use it.
Actually: The verification happens during training only. At inference time the model is just producing its best guess, and it can still be wrong.

FAQ

What does RLVR stand for?
Reinforcement Learning with Verifiable Rewards: reinforcement learning where a program automatically checks the outcome to produce the reward.
What is the difference between RLVR and RLHF?
RLHF rewards come from a model trained on human preference judgments; RLVR rewards come from an objective automatic check such as passing tests, so it suits math and code but not subjective qualities.
Why does RLVR matter for coding agents?
Code has a natural verifier in tests and compilers, so RLVR-style training is a common way to improve a model's ability to solve programming tasks.

Last checked: 2026-09-20

Related terms

Large Reasoning Model (LRM)

A large reasoning model (LRM) is a language model trained to work through a problem step by step in a long internal reasoning trace before giving its final answer.

Test-Time Compute

Test-time compute is computation spent while a model answers a query — rather than during training — used to improve the result through longer chains of reasoning, multiple sampled attempts, search over options, or self-checking.

AI Agent Evaluation

Agent evaluation is the practice of testing an AI agent's complete multi-step behavior against a set of tasks, rather than grading a single output in isolation.

Chain-of-Thought (CoT)

Chain-of-thought is a prompting technique where a model is asked to write out its intermediate reasoning steps before giving a final answer, rather than producing the answer directly.