elloworld
  • About
Categories
All (2)
mechinterp (2)

You don't need error nodes, you need better features

Scaling replacement-aware SAE training to Gemma-2-2B.

Jul 27, 2026
Evan Lloyd
13,105 words

Replacement-aware SAE Training

In which a term for the feature-space error of the next layer's SAE is added to the standard SAE loss function.

Jan 12, 2026
Evan Lloyd
4,441 words
No matching items