Resolve the statistical significance of AD-WM's Scene performance gain

Determine whether AD-WM's improvement over the matched LeWM reproduction on the Scene benchmark is statistically reliable across additional paired seeds and evaluations.

Background

AD-WM outperforms the matched LeWM reproduction on the Scene environment, with success increasing from 35.5% to 39.5%. However, the reported comparison uses only three matched seeds, and the paired seed test does not establish statistical significance. The authors therefore leave unresolved whether the observed Scene improvement reflects a robust performance gain or variation across seeds.