RLVR

When an AI is trained, it needs feedback: Was this answer good or bad? Normally, humans evaluate these answers, which is expensive and slow. RLVR takes a different approach. It trains the model only on tasks where the outcome can be checked by machine — for example, math problems where 42 is either right or wrong. The model then automatically receives feedback on whether it solved the task or not, and adjusts accordingly. Human evaluators are not needed for this.

Why RLVR makes models better at calculating

Most large language models — that is, systems like ChatGPT — are good at sounding fluent. Calculating logically correctly, on the other hand, is harder for them. This is due to how they were trained: they learned to produce probable word sequences, not necessarily correct conclusions.

This is where RLVR comes in. Because the model receives direct, unambiguous feedback, it doesn’t just memorize patterns but develops strategies that actually lead to the correct result. Researchers have observed that models begin, on their own initiative, to check intermediate steps, discard solution paths, and start over — similar to a student double-checking a problem.

A well-known example is the model DeepSeek-R1, which caused a stir in early 2025. It was trained with RLVR on mathematics and programming tasks and achieved performance comparable to significantly more expensive Western models. This showed that RLVR is not just a nuance for experts, but a genuine lever for training efficiency.

Verifiable feedback as a training engine

The term “verifiable” is the decisive part of the name. The principle originates from the field of reinforcement learning. There, a system learns through trial and error: it acts, receives an evaluation back, and acts better next time. Classic examples are AIs that learn to play chess or video games — there, winning or losing is the evaluation.

RLVR transfers this principle to language models. The evaluation — referred to in technical jargon as a “reward” — is not a human’s gut feeling, but an automatic check: Did the model deliver the correct number? Did the generated program code actually compile and produce the correct result? If yes, there is positive feedback; if no, negative feedback.

The method is so attractive because it can be scaled. A team of ten human evaluators might be able to check a hundred answers per hour. An automatic checker achieves the same in milliseconds and can accompany millions of training steps. The limit lies where answers are ambiguous or hard to verify — open-ended text tasks, ethical questions, and creative writing are hardly suitable for RLVR.

RLVR in products and research news

Anyone reading news articles about “reasoning models” will almost always encounter RLVR in the background. Models like OpenAI’s o1 or o3, Google's Gemini Thinking, and the aforementioned DeepSeek-R1 use similar approaches to draw better conclusions. The term “reasoning” refers to a model’s ability to solve problems through multiple mental steps — and RLVR is one of the most important methods for training this ability.

In everyday life, one rarely encounters RLVR directly by name. But when an AI assistant solves a math problem and shows its calculation path or corrects itself, there is a high probability that RLVR training is behind it. Programming tools like GitHub Copilot also benefit from similar techniques: code can be automatically executed and tested, making it an ideal training domain for RLVR.

A common misconception: RLVR makes models generally smarter. In fact, it only improves the areas where verifiable answers exist. For everything that cannot be clearly classified as right or wrong, other methods are still needed.