multimodalfactbullishA training-free mechanistic interpretability method can provide interpretable and effective robustness against typographic attacksComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02494v1