Summary

Ordered SAEs are more steerable than top- SAEs (1). This was tested on antibody J genes.

1.
Haque R, Turnbull OM, Parsan A, Parsan N, Yang JJ, Deane CM. Mechanistic Interpretability of Antibody Language Models Using SAEs. 2025; Available from: https://arxiv.org/abs/2512.05794