AI Paper Atlas

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

STAR: Test-Time Adaptation Can Enhance Universal Prompt Learning for Vision-Language Models

Yiwei Fu, Hui Wan, Xiao Luo, Minghua Deng

CVPR 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 test-time adaptation (M3) on top of prompt-tuned model (ablation)STAR_L · iNaturalist+SUN averageFPR95: 38.36 -> 14.55Improved
use Fisher-score adaptive soft gating instead of fixed threshold 0.6 log K (ablation)STAR_L · iNaturalist+SUN averageFPR95: 83.02 -> 14.55Improved
use conjugate uncertainty-aware loss instead of softmax entropy minimization (ablation)STAR_L · iNaturalist+SUN averageFPR95: 87.24 -> 14.55Improved
add STAR test-time adaptation to 1-shot LoCoOpLoCoOp_G · ImageNet-1k ID, OOD average (1-shot)FPR95: 33.08 -> 27.03Improved
add STAR test-time adaptation to 16-shot SCTSCT_G · ImageNet-1k ID, OOD average (16-shot)FPR95: 26.62 -> 23.17Improved
add STAR to LoCoOp on TextureLoCoOp_G (1-shot) · Texture (1-shot)FPR95: 50.37 -> 52.25Hurt
add STAR on iNaturalist harmonic accuracyLoCoOp (1-shot) · iNaturalist (1-shot)harmonic accuracy ACC_H: 42.72 -> 41.36Hurt
add STAR on harmonic accuracy averageLoCoOp (16-shot) · ImageNet-1k ID, OOD average (16-shot)harmonic accuracy ACC_H: 39.85 -> 46.04Improved
decrease adaptation batch size from 512 to 32 (ablation)STAR · iNaturalistFPR95Improved
change batch size (ablation)STAR · SUNFPR95No clear effect
switch image encoder from ViT-B/16 to ViT-B/32 (ablation)STAR · iNaturalist and SUNFPR95 / AUROCNo clear effect

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