Paulius Rauba (@PauliusRauba) profile picture

Large-scale post-training of long-horizon agentic models at Cursive. PhD @ Cambridge.

London
Joined August 2020
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
1
1
71