$K$-MSHC: Unmasking Minimally Sufficient Head Circuits in Large Language Models with Experiments on Syntactic Classification Tasks

Pratim Chowdhary, Peter Chin, Deepernab Chakrabarty

Understanding which neural components drive specific capabilities in mid-sized language models ($\le$10B parameters) remains a key challenge. We introduce the $(\mathbf{K},\epsilon )$-Minimu…