AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

The Extra Tokens Matter: Disentangled Representation Learning with Vision Transformers

Maofeng Tang, Hairong Qi

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining

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
add minimum volume constraint (MVC) to factor tokens (ablation)XTRA · ImageNet-1KKNN accuracy (%): 72.4 -> 84.2Improved
add factor representation without MVC (ablation)frozen teacher baseline · ImageNet-1KKNN accuracy (%): 76.0 -> 72.4Hurt
freeze DINOv2 teacher (ablation)baseline · ImageNet-1KKNN accuracy (%): 76.1 -> 76.0No clear effect
use multi-stage aggregation (at least 1 stage) instead of none (ablation)XTRA · ImageNet-1KKNN accuracy (%): 13.9 -> 84.2 (final 2-stage model)Improved
add third aggregation stage over two (ablation)XTRA · ImageNet-1KKNN accuracy (%): KNN increases 0.9%, linear probing increases 0.2%No clear effect
use hard (one-hot) assignment instead of soft softmax (ablation)XTRA with MVC · ImageNet-1KSEPIN@1: 1.52 -> 3.95Improved
replace MVC with VICReg-L patch decorrelation (ablation)XTRA · PartImageNet valPart mIoU: 46.8 -> 43.8Hurt
unfreeze more student blocks (final block to all) (ablation)XTRA · ImageNet-1KKNN accuracy (%): 78.3 -> 84.2Improved
train XTRA with factor tokens on top of DINOv2 teacherDINOv2 · ImageNet-CmCE: 41.3 -> 38.2Improved
apply XTRA on different teachers (ablation)MAE / CLIP / DINOv2 · ImageNet-1KKNN accuracy (%): 65.2 -> 68.4 (MAE); 75.8 -> 79.3 (CLIP)Improved
add XTRA without pretrained teacher (from scratch)DINOv2 ViT-B · ImageNet-1K (from scratch)KNN accuracy (%): 76.9 -> 81.9Improved
add register tokens without regularization (DINOv2+)DINOv2 · ImageNet-1kSEPIN@1: 0.47 -> 0.42No clear effect

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