GRGrad@Grad62304977TECH
Tbh even if PPO is the right thing to do for compaction/long horizon, the justification here is pretty bad
Can easily just calculate the mean on a per rollout level not per sample level
Also the paper doesn't ablate it at all against GRPO which is weird https://x.com/Grad62304977/status/2074357111086285291/photo/1