Seems like everyone is converging on masked importance sampling, score centering, and top-p sampling with mask replay as their stable async RL rollout recipe on top of any additional stuff that is infra-specific around rewards/environments and reward hacking