RLVR, short for reinforcement learning with verifiable rewards, trains a language model by rewarding answers that a program can check as correct, such as a maths result.
In plain words
Classic RLHF scores answers with a reward model trained on human preferences. RLVR swaps that reward model for a simple checker: the answer is right or it is not. It works where correctness can be tested, like maths problems or instructions with clear rules. Ai2 reported it improved maths scores such as GSM8K without hurting other skills, and reasoning models such as DeepSeek-R1 trained with RL do especially well on checkable tasks like maths and coding.
A quick example
A model solves a word problem and ends with the number 42. A script compares 42 with the known answer. If they match, the model gets a reward and is nudged to reason that way more often; if not, it gets nothing.
Where you'll meet it
- Training reasoning models on maths and code.
- Post-training open models such as AI2's Tulu 3.
Related
Sources
- PaperTulu 3: Pushing Frontiers in Open Language Model Post-Training, Lambert et al. (arXiv) · read 29 Sept 2026
- DocsTulu 3: The next era in open post-training, Allen Institute for AI · read 29 Sept 2026
- PaperDeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, DeepSeek-AI (arXiv) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.