Denoised Mean Difference
A variant of linear probing that extracts a linear direction corresponding to a target concept in a model's activation space by computing the difference between the mean activations of a target class and a control class, followed by denoising. It is commonly applied to extract representations of abstract concepts such as emotion, honesty, and harmfulness.
Timeline (last 90 days)
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.
Denoised mean difference is a variant of the linear probing technique that eliminates noise unrelated to the target concept by averaging or orthogonalizing results across multiple control conditions.
All Facts (2)
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.
50%UnverifiedDenoised mean difference is a variant of the linear probing technique that eliminates noise unrelated to the target concept by averaging or orthogonalizing results across multiple control conditions.
50%