AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

Towards Fine-Grained Robustness: Attention-Guided Test-Time Prompt Tuning for Vision-Language Models

Jia-Wei Hai, Yijun Wang, Xiu-Shen Wei

ICML 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Test-time adaptation of VLMs · 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 attention refinement, augmentation and TV ensemble together vs TPT-EnsembleA-TPT · 8 fine-grained datasets, ViT-B/16, adversarialadversarial accuracy: 31.8 -> 45.7Improved
add token-gradient attention refinement (A-Refine) alone (ablation)A-TPT · 8 fine-grained datasets, ViT-B/16adversarial and clean accuracy: 31.8 -> 31.8No clear effect
add attention-guided augmentation and TV ensemble without A-Refine (ablation)A-TPT · 8 fine-grained datasets, ViT-B/16, adversarialadversarial accuracy: 31.8 -> 32.3No clear effect
add TV-based ensemble on top of A-Refine and A-Aug (ablation)A-TPT · 8 fine-grained datasets, ViT-B/16, adversarialadversarial accuracy: 41.6 -> 45.7Improved
add attention-guided augmentation on top of A-Refine and A-TV (ablation)A-TPT · 8 fine-grained datasets, ViT-B/16, adversarialadversarial accuracy: 38.1 -> 45.7Improved
reduce number of views from 64 to 16 (ablation)A-TPT · Pets, ViT-B/16, adversarialaccuracy and time per sample: 70.5 -> 69.8; 2.71 s -> 0.53 sMixed
use A-TPT instead of R-TPTA-TPT · 8 fine-grained datasets ViT-B/32 adversarialadversarial accuracy: 35.3 -> 40.3Improved
use A-TPT instead of training-time defensesA-TPT · 8 fine-grained, ViT-B/32clean and adversarial accuracy: adv 40.3 vs FARE 40.3; clean 58.6 vs 42.2Improved
A-TPT under CW, DeepFool, FGSM attacks vs R-TPTA-TPT · Flower, DTD, ViT-B/32adversarial accuracy: Flower 51.8 -> 55.4; DTD 34.2 -> 39.9Improved
A-TPT on Aircraft clean accuracy vs CLIPA-TPT · Aircraft, ViT-B/32 cleanaccuracy: 18.1 -> 17.5Hurt

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