Research map / Label-efficient and robust learning / Knowledge distillation
Aggregating Cross-Domain Knowledge via Learnable Tokens for Multi-Teacher Distillation
Wu Ran, Weijia Zhang, ShuYang Pang, JingSheng Liu, Xiaohui Zhang, Yichao Yan, Chao Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the learnable token (ablation) | ACTok ConvNeXt-B student · patch alignment loss | Lf (lower better): 0.29 -> 0.23 | Improved |
| add the cross-space module (ablation) | ACTok · ConvNeXt-B distillation | patch alignment loss: 0.35 -> 0.23 | Improved |
| add the intra-space module (ablation) | ACTok · ConvNeXt-B distillation | alignment losses: Lf 2.60 -> 0.23 | Improved |
| use cosine instead of MSE for DINO objective (ablation) | ACTok ViT-L · fine-grained avg | accuracy: 90.9 -> 92.4 | Improved |
| scale lr of interaction modules by 0.1x during downstream tuning (ablation) | ACTok ViT-L · CUB | accuracy: 89.2 -> 90.8 | Improved |
| use learnable-token head features instead of backbone (ablation) | ACTok ResNet-50 · CUB2011 | accuracy: 73.8 -> 79.7 (CLS head) | Improved |
| ViT teachers instead of CNN teachers for ConvNeXt-B student | ACTok · ImageNet-1K | top-1 accuracy: 83.0 -> 84.1 | Improved |
| distill from MAE/DINO CNN teachers into ConvNeXt-B student | ACTok ConvNeXt-B · ImageNet-1K | top-1 accuracy: 83.8 -> 83.0 vs supervised | Hurt |
| distill into ResNet-50 | ACTok ResNet-50 · ADE20K linear probing | mIoU: 12.4 -> 20.5 | Improved |
| distill into ResNet-50 | ACTok ResNet-50 · ImageNet-1K | top-1 accuracy: 80.9 -> 78.0 | Hurt |
| ACTok ViT-B vs MAE teacher | ACTok ViT-B · ADE20K linear probing | mIoU: 22.2 -> 46.5 | Improved |
| ACTok ViT-B vs DINOv3 teacher | ACTok ViT-B · ADE20K linear probing | mIoU: 50.9 -> 46.5 | Hurt |
Closest papers by meaning
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.89)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.89)
- Beyond Next-Token Alignment: Distilling Multimodal Large Language Models via Token Interactions (ICML 2026 · similarity 0.88)
- Multi-Teacher Knowledge Distillation with Reinforcement Learning for Visual Recognition (AAAI 2025 · similarity 0.88)
- Learning Task-Agnostic Representations through Multi-Teacher Distillation (NeurIPS 2025 · similarity 0.87)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
