Interpretability paper critique: denoising may project out sub-components beyond noise
dcshiller · x · 2026-10-09
dcshiller raises two methodological concerns about an emotion/pain-vector paper: (a) the effect of subtracting control vectors is unclear — removing something that isn't there effectively adds its inverse; (b) the denoising step projects out the control group's principal components, which may also remove sub-groups like the body component, going beyond denoising and complicating interpretation of the subtraction.
Related event: Researchers Challenge the 'Pain Axis' Paper as AI Welfare Debate Deepens(19 posts)→