Frontier RLVR Post-Training From Scratch: Critic-Free GRPO/DAPO, DoRA, 2D Ring-Attention & Step-Level PRMs
Clean-room PyTorch 2.x & C++20 reference engineering (VeritasRL): 8.0× GRPO/DAPO Pass@1 lift, 563× lower MSE with DoRA vs LoRA, <1.5e-6 2D Ring-Attention autograd parity, and 100% Best-of-6 PRM scaling.