AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention

Weichen Zhou, Yawen Zou, Chunzhi Gu, Ran Dong, Haoran Xie, Chao Zhang

ECCV 2026 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning · 0 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
add point-wise MLP to the linear probeDINOv2-L · NYU Depth V2 (frozen ViT-L, multi-layer features)SA-delta1: 0.9157 -> 0.9325Improved
add global DPT decoder over the MLP probeDINOv2-L · NYU Depth V2 (frozen ViT-L, multi-layer features)SA-delta1: 0.9325 -> 0.9483Improved
add point-wise MLP to the linear probeMAE-L · NYU Depth V2 (frozen ViT-L, multi-layer features)SA-delta1: 0.6033 -> 0.6390Improved
add global DPT decoder over the linear probeMAE-L · NYU Depth V2 (frozen ViT-L, multi-layer features)SA-delta1: 0.6033 -> 0.7022Improved
add DPT decoder over the linear probeiBOT-L · NYU Depth V2 (frozen ViT-L, multi-layer features)SA-delta1: 0.8198 -> 0.8524Improved
restrict probe to random or residual control subspaces instead of the top-k task-aligned subspace (ablation)linear probe · NYU Depth V2SA-delta1: collapses below 0.18Hurt
increase subspace rank to k=32 (ablation)MAE-L linear probe · NYU Depth V2recovery of linear baseline: >98% of linear potential at k=32Improved
increase subspace rank to k>=64 (ablation)DINOv2-L linear probe · NYU Depth V2recovery of linear baseline: converges at k >= 64Improved
extract layer-24 final readout instead of intermediate layersDINOv2-L · NYU Depth V2geometric energy contribution: 35.8 / 36.7 at l12/l18 -> 10.3 at l24Hurt
choose layer 18, 21 or 24 as readout for normals, depth, segmentationDINOv2-L · NYUv2 linear probespeak layer by task: normals peak layer 18, depth layer 21, segmentation layer 24Mixed
use low rank k<=16 subspace across random probe seeds (ablation)DINOv2-L · NYU Depth V2subspace similarity: >0.93 at k <= 16Improved

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