Beyond Score Centering: What Else Can We Optimize in Async RL?
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Score centering (SC) addresses instability in asynchronous reinforcement learning for LLMs, where the trainer learns from trajectories sampled by a policy several steps old. The post shows that the policy-gradient norm grows roughly linearly with trajectory age, because an off-policy drift term appears.