intervention-scaling-monotonicity-king-to-queen

IN premise — summaries/2026/08/24/park-2023-linear-representation-s2-king-queen-queen-queen-queen.md

Created 2026-08-24T17:11:02+00:00

In the Park et al. (2023) intervention experiment on context 'Long live the ', adding α·v_W (male⇒female direction) with α increasing from 0 to 0.4 shifts 'king' from rank #1 to exiting the top-5 entirely, while 'queen' reaches ranks 1–2 by α = 0.3, demonstrating monotonic linear control over concept activation in logit space.

Summary

Pushing a language model's internal math a little further in the "male-to-female" direction reliably and proportionally swaps its top word prediction from "king" to "queen," with no surprises or reversals along the way. This matters because it shows conceptual meaning in the model lives in a predictable geometric space you can steer with a simple, linear nudge, rather than requiring brute-force retraining.