Research map / Architectures and efficiency / Attention and transformers
Representation Shift: Unifying Token Compression with FlashAttention
Joonmyung Choi, Sanghyeok Lee, Byungoh Ko, Eunseo Kim, Jihyung Kil, Hyunwoo J. Kim
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 |
|---|---|---|---|
| replace attention-score importance with representation shift (plus FlashAttention) in token pruning | UMT-L · ActivityNet R@1 | R@1: 53.2 -> 62.9 | Improved |
| replace attention-score importance with representation shift | UMT-B · MSVD R@1 | R@1: 60.3 -> 57.7 | Hurt |
| replace attention-score importance with representation shift | DeiT-S · ImageNet-1K | top-1 accuracy: 72.1 -> 77.8 | Improved |
| replace attention-score importance with representation shift | UMT-L · MSVD-QA | accuracy: 54.2 -> 54.9 | Improved |
| swap vid-TLDR attention importance for representation shift with FlashAttention | vid-TLDR UMT-B · MSRVTT | throughput (vid/s): 43.6 -> 131.4 | Improved |
| use representation shift through MLP only instead of attention or whole block (ablation) | DeiT-S / UMT-B · ImageNet-1K / MSRVTT | accuracy | Improved |
| use L2 distance instead of L1 or cosine (ablation) | DeiT-S / UMT-B · ImageNet-1K / MSRVTT | accuracy | Improved |
| keep top-50% tokens by representation shift instead of bottom-50% (ablation) | DeiT-S · ImageNet-1K | top-1 accuracy: 51.7 -> 78.0 | Improved |
| use line-wise instead of token-wise pruning (ablation) | ResNet-50 · ImageNet-1K | top-1 accuracy: 75.9 -> 76.4 | Improved |
| replace ToP-ViM importance with representation shift | ViM-T · ImageNet-1K | top-1 accuracy: 75.1 -> 75.5 | Improved |
| prune tokens with representation shift on top of FlashAttention | DeiT-S · ImageNet-1K | top-1 accuracy: 79.8 -> 77.8 | Hurt |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.88)
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.88)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.88)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.88)
- TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model (CVPR 2026 · 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.
