1. INP-Former, CVPR 25
2. Dinomaly, cvpr 25 (https://arxiv.org/abs/2405.14325): The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection.
   1. 在 GitHub 上，项目 README 中称其为 “the first multi-class UAD model that can compete with single-class SOTAs”
   2. minimalistic / simple design: only Transformer (Attention + MLP), no much dependency on complex modules or tricks.
   3. reconstruction-based
      1. 但它并不是典型的自编码器式重构，而是对重构机制做了“弱化”、控制重构能力、防止过拟合，以使其在多类统一模型的场景下更稳健。⇒
   4. loose reconstruction instead of tight reconstruction (traditional):
      1. do not force layer-to-layer and point-to-point reconstruction
         1. layer-to-layer ⇒ use high/coarse level features only
         2. point-to-point ⇒ cosine similarity instead of L2 loss
      2. Noisy Bottleneck
         1. Dropouts do all the noise injection tricks
      3. Selective masking / skipping of easy regions in training
   5. ShortcutBreaker: Low-Rank Noisy Bottleneck with Global Perturbation Attention for Multi-Class Unsupervised Anomaly Detection, 25
4. CostFilter-AD, ICML 25 (https://arxiv.org/pdf/2505.01476): Enhancing Anomaly Detection through Matching Cost Filtering
5. Unified Unsupervised Anomaly Detection via Matching Cost Filtering, 25
   1. https://arxiv.org/pdf/2510.03363
   2. builds upon CostFilter-AD [42] with four advances.
6. Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt, AAAI24 (https://arxiv.org/abs/2401.01010)
   1. 新交叉方向：Continual Learning（持续学习） + Anomaly Detection（异常检测） + Prompt Learning（提示学习）
8. Generalist Multi-Class Anomaly Detection via Distillation to Two Heterogeneous Student Networks, 25
9. Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs, 25
   1. 做LLM based anomaly应该看看这个