Figureure 1 · : An overview of the proposed methodFig. 1: An overview of the proposed method. a) shows the sampling-based mechanistic interpretability (the Stochastic Lottery) for lexical circuit mining. Latent samples (denoted by purple dashed arrows) on the hypothesized concept basis reveal distinct attribution patterns. b) illustrates the overall sampling and circuit-mining process using gradient-based attribution. c) shows the mechanistic intervention process. To improve TA robustness, we intervene in vulnerable attention modules via attention reweighting or zero-ablation.这张图概括 Towards Robustness against Typographic Attack 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。Figureure 8 · : Mean nTAS of Attention Heads across ViT layers.Fig. 8: Mean nTAS of Attention Heads across ViT layers.这张可视化用来解释 Towards Robustness against Typographic Attack 学到的中间表征或对齐关系。重点看它是否支持正文里的机制判断。
