AI Paper Atlas

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

ICCV 2025 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 4 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace attention-score importance with representation shift (plus FlashAttention) in token pruningUMT-L · ActivityNet R@1R@1: 53.2 -> 62.9Improved
replace attention-score importance with representation shiftUMT-B · MSVD R@1R@1: 60.3 -> 57.7Hurt
replace attention-score importance with representation shiftDeiT-S · ImageNet-1Ktop-1 accuracy: 72.1 -> 77.8Improved
replace attention-score importance with representation shiftUMT-L · MSVD-QAaccuracy: 54.2 -> 54.9Improved
swap vid-TLDR attention importance for representation shift with FlashAttentionvid-TLDR UMT-B · MSRVTTthroughput (vid/s): 43.6 -> 131.4Improved
use representation shift through MLP only instead of attention or whole block (ablation)DeiT-S / UMT-B · ImageNet-1K / MSRVTTaccuracyImproved
use L2 distance instead of L1 or cosine (ablation)DeiT-S / UMT-B · ImageNet-1K / MSRVTTaccuracyImproved
keep top-50% tokens by representation shift instead of bottom-50% (ablation)DeiT-S · ImageNet-1Ktop-1 accuracy: 51.7 -> 78.0Improved
use line-wise instead of token-wise pruning (ablation)ResNet-50 · ImageNet-1Ktop-1 accuracy: 75.9 -> 76.4Improved
replace ToP-ViM importance with representation shiftViM-T · ImageNet-1Ktop-1 accuracy: 75.1 -> 75.5Improved
prune tokens with representation shift on top of FlashAttentionDeiT-S · ImageNet-1Ktop-1 accuracy: 79.8 -> 77.8Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy