Preprint2026
Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations
Pere Martra, Eugenio Martínez Cámara, Alfonso Ureña López
Fairness Pruning introduces a lightweight structural intervention that locates demographic bias in GLU-MLP layers by zeroing at most 40 neurons, achieving 99.49% capability retention while destabilizing bias bidirectionally.
0Jul 30, 2026Large Language ModelsPrompting
arXiv