Breaking entropy bounds: Accelerating RL training via MTP with rejection sampling. Given that RL training is limited by rollouts, and MTP is used to accelerate them, its acceptance rate would still collapse as the entropy of the

→ View original post on X — @askalphaxiv