Figureure 3 · Class accuracy versus log-scaled efective information for retained preFig. 3. Class accuracy versus log-scaled efective information for retained prefix budgets across tuned architectures and regularizer-matched controls. Arrows indicate directional changes of the respective weights in the control setting. Single arrows indicate both variants, two arrows indicate ITD/TD changes.这张图概括 Semantic Allocation in Ordered Bottlenecks 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 2 · Budget-wise linear-probe accuracy for class and coarse taxonomy levelsFig. 2. Budget-wise linear-probe accuracy for class and coarse taxonomy levels. Shading decomposes readout quality by utilized tokens up to the highest-performing prefix of the respective model; for PRIOR, a prefix corresponds to the first k concatenated levels.这张图/表用于判断 Semantic Allocation in Ordered Bottlenecks 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。