Per-Problem Base Anchoring (PBA)

A new arXiv study (2607.20543v1) finds that reinforcement learning with verifiable rewards (RLVR) can improve one-sample accuracy while causing pass@k inversion, where the trained policy solves fewer …