AI Paper Atlas

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

D$^2$O: A Dual Debiasing Operator for Training-Free Test-Time Adaptation of Vision–Language Models

Yihong Luo, Wenwu He, Dong Liang, Yihang Zhou, Zhuo-Xu Cui

ICML 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Open-vocabulary alignment from VLMs

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 D2O to ADAPTADAPT · PUG (online)top-1 accuracy (avg): 51.87 -> 58.28 (+6.41)Improved
add D2O to ADAPTADAPT · PUG (transductive)top-1 accuracy (avg): 52.85 -> 56.85 (+4.00)Improved
add D2O to TDATDA · PUG (online)top-1 accuracy (avg): 51.50 -> 52.22 (+0.72)Improved
add D2O to ADAPTADAPT · ImageNet OOD avg (online)top-1 accuracy: 65.44 -> 65.92 (+0.48)Improved
add D2O to ADAPTADAPT · ImageNet-C (online)top-1 accuracy (avg): 28.56 -> 29.15Improved
add D2O to ADAPTADAPT · Fine-grained 10 datasets (online)top-1 accuracy (avg): 70.76 -> 71.12Improved
add prior debiasing (ablation)ADAPT with content evidence only · PUG+ImageNet (online)top-1 accuracy: 58.70 -> 62.10Improved
add retrieval-oriented content evidence (ablation)ADAPT baseline · PUG+ImageNet (online)top-1 accuracy: 58.66 -> 58.70No clear effect
add retrieval-oriented content evidence (ablation)ADAPT with prior debias · PUG+ImageNet (online)top-1 accuracy: 62.01 -> 62.10No clear effect
add content features for retrieval (ablation)raw-feature retrieval · PUG-ImageNet Camera factorTop-5 retrieval purity: +2.5%Improved
use routed bias instead of single global EMA bias (ablation)D2O · Mixed-style PUG streams (online)top-1 accuracy: 47.56 -> 42.78 (-4.78 for G-EMA)Improved
increase number of routing clusters to 32 or more (ablation)D2O+ADAPT · PUG-ImageNet (online)top-1 accuracyHurt

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