Research map / Architectures and efficiency / Attention and transformers
Plug-and-Play Attention Linearization for Pretrained Transformers
Kenan Kassab, Alexey Kashevnik, Ammar Ali, Stamatios Lefkimmiatis
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 Tunable Taylor Expansion (TTE) on top of JPN (ablation) | LATTE (CLIP) · MSCOCO Captions retrieval (67% attention layers linearized) | text recall: 65.1 -> 77.1 | Improved |
| add Joint Post-Squaring Normalization (JPN) on top of TTE (ablation) | LATTE (CLIP) · MSCOCO Captions retrieval (67% attention layers linearized) | text recall: 71.5 -> 77.1 | Improved |
| add JPN and TTE to QT-ViT (ablation) | LATTE (CLIP) · MSCOCO Captions retrieval (67% attention layers linearized) | text recall: 42 -> 77.1 | Improved |
| add TTE only to QT-ViT (ablation) | LATTE (SigLip) · MSCOCO Captions retrieval (67% attention layers linearized) | text recall: 82.3 -> 86.3 | Improved |
| add JPN only to QT-ViT (ablation) | LATTE (SigLip) · MSCOCO Captions retrieval (67% attention layers linearized) | text recall: 82.3 -> 83.2 | Improved |
| LATTE instead of QT-ViT at 67% linearization | RT-DETR · COCO detection | mAP@0.5:0.95: 40.6 -> 46.7 | Improved |
| LATTE instead of QT-ViT at 67% linearization | DETR · COCO detection | mAP@0.5: 52.5 -> 59.2 | Improved |
| LATTE instead of QT-ViT at 67% linearization | CLIP · MSCOCO Captions retrieval (67% attention layers linearized) | image recall: 31.2 -> 59.7 | Improved |
| LATTE instead of QT-ViT at 75% linearization | CLIP · SugarCrepe (75% linearized) | text recall: 5.30 -> 95.91 | Improved |
| LATTE instead of QT-ViT at 54% linearization | CLIP · CIFAR-10 (54% linearized) | top-1 accuracy: 86.88 -> 90.56 | Improved |
| vary calibration subsample from COCO training set (ablation) | LATTE (RT-DETR) · COCO detection | mAP50:95 | No clear effect |
| raise layer-selection lambda from 0.1 to 5 or more (ablation) | LATTE (RT-DETR) · RT-DETR COCO (67% linearized) | mAP50:95: 0.44 -> 0.459 | Improved |
Closest papers by meaning
- The Hedgehog & the Porcupine: Expressive Linear Attentions with Softmax Mimicry (ICLR 2024 · similarity 0.88)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.88)
- STILL: Selecting Tokens for Intra-Layer Hybrid Attention to Linearize LLMs (NeurIPS 2026 · similarity 0.88)
- MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention (NeurIPS 2025 · similarity 0.87)
- QT-ViT: Improving Linear Attention in ViT with Quadratic Taylor Expansion (NeurIPS 2024 · similarity 0.87)
- PolaFormer: Polarity-aware Linear Attention for Vision Transformers (ICLR 2025 · 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.
