Tensorwire
Products & tools · first seen 7 Sep, updated 7 Sep

Recurrence Is Not Enough: Causally Validating Multilingual SAE Translation Features in Gemma 2 and 3

1 outlet Gemma

arXiv:2609.04808v1 Announce Type: cross Abstract: Sparse autoencoder (SAE) features are increasingly used to explain and steer language-model behavior, but it remains unclear whether a feature found in one language context plays the same ca…

Summary from arXiv cs.AI.

Coverage 1 article · 1 outlet

  1. arXiv cs.AI
    Recurrence Is Not Enough: Causally Validating Multilingual SAE Translation Features in Gemma 2 and 3