AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Model Merging in the Essential Subspace

Longhua Li, Lei Qi, Qi Tian, Xin Geng

CVPR 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 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 SVD decomposition with essential subspace decomposition (ESD) (ablation)TSV-M-style merging · 8-task merging, CLIP ViT-B/16average accuracy (%): 89.6 -> 90.9Improved
replace SVD decomposition with ESD (ablation)TSV-M-style merging · 20-task merging, CLIP ViT-L/14average accuracy (%): 88.1 -> 89.5Improved
add polarized scaling (three levels) on ESD (ablation)ESD merging · 20-task merging, CLIP ViT-B/16average accuracy (%): 82.8 -> 84.9Improved
add polarized scaling on SVD-based merging (ablation)TSV-M · 20-task merging, CLIP ViT-B/16average accuracy (%): 80.6 -> 82.1Improved
add inter-task and inter-dimension scaling on top of layer scaling (ablation)ESM · 8-task merging, CLIP ViT-B/16average accuracy (%): 91.4 -> 91.8Improved
reverse polarized scaling (reciprocal factors) (ablation)ESM · 8-task merging, CLIP ViT-B/32average accuracy (%): 86.7 -> 82.9Hurt
suppress noise only (factors < 1) (ablation)polarized scaling · 8-task merging, CLIP ViT-B/32average accuracy (%): 86.7 -> 87.4Improved
amplify signal only (factors > 1) (ablation)polarized scaling · 8-task merging, CLIP ViT-B/32average accuracy (%): 86.7 -> 87.7Improved
use out-of-distribution ImageNet-1k samples as proxy set instead of in-distribution (ablation)ESM · 8-task merging, CLIP ViT-B/32average accuracy (%): 88.4 -> 88.3No clear effect
use single-class proxy samples (ablation)ESM · 8-task merging, CLIP ViT-B/32average accuracy (%): 88.4 -> 88.4No clear effect
use even a single proxy sample per task (ablation)ESM · 8-task merging, CLIP ViT-B/32average accuracy (%)Improved
replace TSV-M/Iso-CTS with ESMIso-CTS · 20-task merging, CLIP ViT-B/32average accuracy (%): 78.1 -> 81.3Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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