X · @teortaxesTex
· X / Twitter
most of the gain here comes from Direct Double-Sided Importance Sampling (DIS). Vanilla GRPO is long obsolete, but it's surprising how many tweaks hav…
most of the gain here comes from Direct Double-Sided Importance Sampling (DIS). Vanilla GRPO is long obsolete, but it's surprising how many tweaks have been discovered to prolong its relevancestill, seems like SAO has an even higher ceilingjietang: Our Single-rollout Asynchronous Optimization (SAO), is able to train st