We introduce a RL framework to train LLM's reasoning and self-verification ability simultaneously.

How do models generalize to unseen constraints for precise instruction following?

We only need one example for RLVR on LLMs to achieve significant improvement on math tasks