The study uses a denoised difference-in-means method to extract linear pain directions from 25 open-weight models spanning five model families with parameter scales ranging from 2B to 72B.