Figureure 2 · Uniform linear prior suppression can hurtFigure 2. Uniform linear prior suppression can hurt. VCD (vs. vanilla) shows consistent performance drops on POPE across splits, including standard cases where priors align with visual evidence.这张图/表用于判断 Mitigating Manifold Departure 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致,而不是只看单个最高分。Figureure 4 · Overview of the proposed Manifold-Guided Adaptive Projection (MGAP)Figure 4. Overview of the proposed Manifold-Guided Adaptive Projection (MGAP). The language-prior subspace is constructed offline using unlabeled blind inputs, while decoding-time intervention is performed online via geometry-aware adaptive projection.这张图概括 Mitigating Manifold Departure 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。