notesum.ai

Published at November 13

Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers

cs.CL
cs.AI

Released Date: November 13, 2024

Authors: Clément Dumas1, Chris Wendler2, Veniamin Veselovsky2, Giovanni Monea3, Robert West2

Aff.: 1ENS Paris-Saclay; 2EPFL Lausanne; 3Cornell Tech

Arxiv: http://arxiv.org/abs/2411.08745v1