REINFORCE Zero · importance-weight deviation D3

With standard kernels, importance-weight deviation peaks at 442× its median

D3 is the per-update mean of w − 1 − log w, which is 0 when every importance weight w is 1. Bars show its peak over the run divided by its median.

DeepSeek-R1-Distill-Qwen-1.5B, two asynchronous runs · REINFORCE Zero, Table 2