Head-level lesion-symptom mapping of picture naming in vision-language models
Samaneh Nemati, Roger Newman‐Norlund, Saeed Ahmadi, Xiang Guan, Kalil Warren, Yong Yang, Srihari Nelakuditi, Chris Rorden, Leonardo Bonilha, Julius Fridriksson
The paper ablates attention heads in vision-language models to study picture naming accuracy, finding that head removal disrupts naming but does not confirm the head computes the behavior.
bioRxiv · Sep 21, 2026 · PreprintNeuroscience