扩散模型因忽略稀有关键结构而失效，新方法通过谱权重修复缺陷。
Generative Myopia: Why Diffusion Models Fail at Structure
- 用有效电阻重构扩散目标，让模型关注稀有但关键的结构
- 在对抗性测试中实现100%连通性，传统方法为0%
- 训练时引入谱先验，推理无额外开销，适合图结构生成任务
图扩散模型（GDMs）优化统计似然，隐式充当频率滤波器，偏好常见子结构而非谱上关键的结构，这种现象称为生成性短视（Generative Myopia）。在图稀疏化等组合任务中，会导致灾难性地移除‘稀有桥接边’——这些边在结构上不可或缺（$R_{ ext{eff}} \approx 1$）但统计上稀少。我们从理论和实证上证明，该失败源于梯度饥渴：优化景观本身抑制了稀有结构信号，使其不可学习，与模型容量无关。为此，我们提出谱加权扩散（Spectrally-Weighted Diffusion），通过有效电阻重新对齐变分目标。实验表明，谱先验可被摊销到训练阶段，推理零开销。该方法消除短视问题，性能媲美最优谱预言机，在标准扩散模型完全失败的对抗性基准上实现100%连通性。
原文摘要 · Abstract (English)
Graph Diffusion Models (GDMs) optimize for statistical likelihood, implicitly acting as \textbf{frequency filters} that favor abundant substructures over spectrally critical ones. We term this phenomenon \textbf{Generative Myopia}. In combinatorial tasks like graph sparsification, this leads to the catastrophic removal of ``rare bridges,'' edges that are structurally mandatory ($R_{\text{eff}} \approx 1$) but statistically scarce. We prove theoretically and empirically that this failure is driven by \textbf{Gradient Starvation}: the optimization landscape itself suppresses rare structural signals, rendering them unlearnable regardless of model capacity. To resolve this, we introduce \textbf{Spectrally-Weighted Diffusion}, which re-aligns the variational objective using Effective Resistance. We demonstrate that spectral priors can be amortized into the training phase with zero inference overhead. Our method eliminates myopia, matching the performance of an optimal Spectral Oracle and achieving \textbf{100\% connectivity} on adversarial benchmarks where standard diffusion fails completely (0\%).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。