search

vgel / repeng

A library for making RepE control vectors
https://vgel.me/posts/representation-engineering/
MIT License
483 stars 39 forks source link

Why only 1 direction during PCA? #54

Open Hellisotherpeople opened 3 days ago

Hellisotherpeople commented 3 days ago

Can't we use many more directions than just 1? If so, how?

thiswillbeyourgithub commented 1 day ago

Not OP but :

  1. The first dim by PCA is the one that exlains the most variance
  2. The activation of the network is a 1 dimensional array anyway