park-2023-intervention-logit-space-addition

IN premise — summaries/2026/08/24/park-2023-linear-representation-s3-queen-king-lady-lady.md

Created 2026-08-24T17:11:03+00:00

The intervention in Park et al. (2023) is a logit-space addition (λ_{W,α}(x) = λ(x) + α·λ̄_W for α ∈ [0,1]), not an activation-space edit; at α=0.1 on context 'Long live the', top-1 shifts from 'king' to 'Queen' and 'king' drops below top-5.

Summary

The Park et al. (2023) method works by adding a small, scaled value to the model's final output scores rather than surgically editing its internal representations. A nudge as mild as 10 percent strength is enough to flip the top prediction from "king" to "Queen" for the prompt "Long live the," which means the model's gender association is shallow and can be overridden with minimal intervention at the output layer.