Reinforcement Learning from Verifiable Rewards

AI and Machine Learning · Language Models · 2025 · also: RLVR · verifiable-rewards.yaml

Replaces the learned reward with a programmatic check — does the proof verify, do the tests pass, is the answer right. Works only where correctness is decidable, and there the signal cannot be gamed.

• supersedescorrects · extends
• classifiesspecializes · part-of
• substitutes forapproximates · alternative-to
• depends onrequires · validates

Colour is the family; a dashed line is the second member of it.

This node

• correctsfixes a defect in RLHFa learned reward can be gamed; a checked answer cannot
• requiresdoes not work without Group Relative Policy Optimizationone scalar per completion, with no critic to fit

References

• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — DeepSeek-AI — 2025 · link