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

AdMiT: Adaptive Multi-Source Tuning in Dynamic Environments

CVPR 2025 · Label-efficient and robust learning › Test-time adaptation · approach: Test-time adaptation of vision-language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AdMiT, which pretrains a bank of frozen-backbone PET modules on source distributions and, at test time, selects and averages top-M modules via kernel mean embedding matching of few-shot unlabeled target batches, followed by sharpness-aware pseudo-label entropy tuning.

Problem

Adapting large pretrained transformers to dynamic, unlabeled, small-batch target distributions at test time on resource-constrained edge devices, without raw source data, routing networks, or heavy hyperparameter tuning, while retaining source knowledge.

Approach

Matches target batches to source PET modules by approximating the target empirical kernel mean embedding as a weighted combination of source KMEs (computed on synthetic data), then forms a new module as a weighted average of the top-M source modules and fine-tunes it with sharpness-aware pseudo-label entropy minimization, with no routing network.

Key results

• On Digits-Five static adaptation, AdMiT reaches 86.8% average accuracy versus 86.1% for the best multi-source baseline CONTRAST and 90.1% for ground-truth tuning.
• On ImageNet-C at batch size 1, AdMiT achieves 58.7% versus 57.9% for CONTRAST and 60.3% for GT-Tuning.
• On CIFAR-100C dynamic adaptation with 15 sequential domains, AdMiT averages 73.3% on source-domain retention versus 72.5% for CONTRAST.

Measured effects reported in the paper

Surprising findings

• Model soup performs worse than most baselines on Digits-Five and ImageNet-C despite being a multi-source averaging method.
• Plain SGD tuning (AdMiT-Plain) sometimes reduces accuracy relative to the zero-shot integrated module, which motivated the sharpness-aware variant.

Limitations

• Assumes the target distribution can be approximated as a linear mixture of source distributions.
• Relies on a public feature extractor (DenseNet201 pretrained on ImageNet) for the shared embedding space.
• Accuracy gap to ground-truth tuning remains large, and adaptation is limited to image classification in the main paper; segmentation is only in the appendix.

Experimental setup

Datasets: Digits-Five ImageNet-C CIFAR-100C

Benchmarks: Digits-Five ImageNet-C CIFAR-100C

Metrics: accuracy

Baselines: TENT BECoTTA SAR GT-Tuning π-tuning SESoM CONTRAST Model soup

Builds on: Visual Prompt Tuning LoRA SAR TENT CONTRAST

Closest papers by meaning

• SNAP: Low-Latency Test-Time Adaptation with Sparse Updates (NeurIPS 2025 · similarity 0.86)
• Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation (NeurIPS 2024 · similarity 0.86)
• RoSA: Rotational Sparse Adaptation for Memory-Efficient Fine-Tuning (NeurIPS 2026 · similarity 0.85)
• Mixture-of-Control: State-Aware Fine-Tuning for Transformer-based Models (NeurIPS 2026 · similarity 0.84)
• WeGeFT: Weight‑Generative Fine‑Tuning for Multi‑Faceted Efficient Adaptation of Large Models (ICML 2025 · similarity 0.84)
• ETHER: Efficient Finetuning of Large-Scale Models with Hyperplane Reflections (ICML 2024 · similarity 0.83)