PIVOT

A method for reinforcement learning from verifiable rewards (RLVR) that inserts teacher correction segments into student trajectories and applies token-level KL supervision. It mitigates the sparse-reward problem and improves the performance of large models on difficult reasoning tasks.