Figureure 2 · Overview of our CRG frameworkFigure 2. Overview of our CRG framework. Step 1 computes modal causal effects for each attention head by separating visual and textual components and measuring their contributions. Step 2 identifies conflicts by comparing the signs of visual and textual effects, distinguishing Agreement vs. Conflict (A/B), and selects the top-k conflicting heads for intervention. Step 3 applies head gating, where Conflict-A heads receive mild gating and Conflict-B heads receive strong gating, reducing language-prior dominance while keeping image-grounded evidence.这张图概括 Causal Route Gating 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 1 · When language priors override visual inputs, the baseline model tends Figure 1. When language priors override visual inputs, the baseline model tends to produce hallucinated predictions (left). To address this, a conflict-aware intervention applies text-route gating to suppress language priors (middle), enabling the model to correctly rely on visual evidence after intervention (right).这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Causal Route Gating 的方法或实验，请结合正文精读段落一起看。
