Excellent work by Jatin Nainai (nainanijatinz.github.io), Bryn Reimer (code.brynmarie.com), and Connor Watts (connorwatts.github.io)
in collaboration with David Jensen (groups.cs.umass.edu/jensen/)

This work builds on recent papers that have used SAEs to uncover interpretable concepts learned by pLMs. We use activation patching to ask: which concepts are causally necessary for the model to perform contact prediction? For our two case studies, we find an explainable, sparse circuit

This was a huge collaborative effort with my talented PhD student @nvlsen.bsky.social, Irene Lepori, Zichen Liu, Shasha Feng, and the labs of @siegristpalmore.bsky.social, Marcos Pires, Joel Freundlich, and Wompil Im.

Thank you Zam! Cool to see the paper featured. I agree with the writer that attention blocks and changes in padding strategies could be useful

Have since learned about this one: bsky.app/starter-pack...