Research map / Architectures and efficiency / Attention and transformers
Block Recurrent Dynamics in Vision Transformers
Mozes Jacobs, Thomas Fel, Richard Hakim, Alessandra Brondetta, Demba E. Ba, T. Anderson Keller
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 autoregressive loss with annealed teacher forcing (ablation) | Raptor k=3 · ImageNet-1k | top-1 accuracy: 3.9 -> 72.7 (+68.8) | Improved |
| add depth scaling (learned layer-index embedding) (ablation) | Raptor k=3 · ImageNet-1k | top-1 accuracy: 72.7 -> 75.2 (+2.5) | Improved |
| up-weight cls token loss in final block (ablation) | Raptor k=3 · ImageNet-1k | top-1 accuracy: 75.2 -> 76.7 (+1.5) | Improved |
| add second stage end-to-end autoregressive training (ablation) | Raptor k=3 · ImageNet-1k | top-1 accuracy: 76.7 -> 82.4 (+5.7) | Improved |
| fine-tune the linear classifier (ablation) | Raptor k=3 · ImageNet-1k | top-1 accuracy: 82.4 -> 83.0 (+0.6) | Improved |
| increase recurrent blocks from k=2 to k=3 | Raptor · ImageNet-1k | top-1 accuracy: 81.2 -> 83.0 | Improved |
| increase recurrent blocks from k=3 to k=4 | Raptor · ImageNet-1k | top-1 accuracy: 83.0 -> 83.2 | No clear effect |
| increase recurrent blocks from k=2 to k=3 | Raptor · ADE20k | mIoU: 39.6 -> 43.0 | Improved |
| increase recurrent blocks from k=2 to k=3 | Raptor · NYUv2 | RMSE (lower better): 0.648 -> 0.618 | Improved |
| increase stochastic depth probability during ViT training (ablation) | ViT-B/14 from scratch · CIFAR-100 | layerwise similarity and Raptor reconstruction fidelity | Improved |
| compare Raptor k=3 to DINOv2 ViT-S | Raptor k=3 · ImageNet-1k | top-1 accuracy: 80.9 -> 83.0 | Improved |
| compare Raptor k=3 to DINOv2 ViT-B | Raptor k=3 · ADE20k | mIoU: 47.5 -> 43.0 | Hurt |
Closest papers by meaning
- Energy Landscape-Aware Vision Transformers: Layerwise Dynamics and Adaptive Task-Specific Training via Hopfield States (NeurIPS 2025 · similarity 0.88)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.87)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.86)
- Dynamical Properties of Tokens in Self-Attention and Effects of Positional Encoding (NeurIPS 2025 · similarity 0.85)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.85)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.85)
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.
