AI Paper Atlas

Research map / Label-efficient and robust learning / Test-time adaptation

Learning to Memorize with Attributive and Associative Memory for Online Test-Time Adaptation of Vision-Language Models

Yuchao Zhang, Hao Wang, Fan Zhang, Qirui Mi, Mengyue Yang, Yisen Wang, Jun Wang, Haoxuan Li, Zhouchen Lin

ICML 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Test-time adaptation of vision-language models

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 prior textual representation T (ablation)A2Memory · ImageNet OOD benchmark (OOD avg) (ViT-B/16)top-1 accuracy: 63.80 -> 65.99Improved
add prior textual representation T (ablation)A2Memory · Cross-dataset generalization (10 datasets, avg) (ViT-B/16)top-1 accuracy: 68.50 -> 71.41Improved
add surrogate visual representation via soft assignment (ablation)A2Memory · ImageNet OOD benchmark (OOD avg) (ViT-B/16)top-1 accuracy: 64.12 -> 65.99Improved
add surrogate visual representation via soft assignment (ablation)A2Memory · Cross-dataset generalization (10 datasets, avg) (ViT-B/16)top-1 accuracy: 69.45 -> 71.41Improved
add class-wise associative memory (ablation)A2Memory · ImageNet OOD benchmark (OOD avg) (ViT-B/16)top-1 accuracy: 63.45 -> 65.99Improved
add class-wise associative memory (ablation)A2Memory · Cross-dataset generalization (10 datasets, avg) (ViT-B/16)top-1 accuracy: 68.95 -> 71.41Improved
use entropy-adaptive retention instead of fixed alpha 0.9 (ablation)A2Memory · ImageNet OOD benchmark (OOD avg) (ViT-B/16)top-1 accuracy: 64.23 -> 65.99Improved
use entropy-adaptive retention instead of random update (ablation)A2Memory · ImageNet OOD benchmark (OOD avg) (ViT-B/16)top-1 accuracy: 41.55 -> 65.99Improved
use entropy-adaptive retention instead of cumulative average (ablation)A2Memory · ImageNet OOD benchmark (OOD avg) (ViT-B/16)top-1 accuracy: 59.90 -> 65.99Improved
fuse memory branch with CLIP zero-shot branch (ablation)A2Memory · ImageNet (ViT-B/16)top-1 accuracy: 71.34 -> 72.21Improved
use GPT-4o attributes instead of hand-crafted templates (ablation)A2Memory · Cross-dataset generalization (10 datasets, avg) 100% stream (ViT-B/16)top-1 accuracy: 71.09 -> 71.53Improved
replace ADAPT with A2MemoryADAPT · Cross-dataset generalization (10 datasets, avg) (ViT-B/16)top-1 accuracy: 70.76 -> 71.41Improved

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