Research map / Label-efficient and robust learning / Test-time adaptation
Mint: A Simple Test-Time Adaptation of Vision-Language Models against Common Corruptions
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.
Identifies variance collapse in CLIP image embeddings under corruptions and proposes Mint, a test-time adaptation method that maximizes pseudo-label inter-class variance via mean and gradient accumulators, working with small batches.
Problem
CLIP's zero-shot accuracy degrades under common image corruptions; existing VLM test-time adaptation methods either ignore image embedding quality or need large batches and lack theory.
Approach
Corruptions cause variance collapse (intra- and inter-class variance both shrink); a theorem shows maximizing pseudo-label inter-class variance over LayerNorm weights suppresses shift components. Mint estimates this objective online with a cumulative mean accumulator (class and global means) and a cumulative gradient accumulator, plus training-free text embedding refinement.
Key results
- Mint reaches 71.0% average accuracy on CIFAR-10-C with ViT-B/32 versus 67.1% for the strongest baseline WATT-S.
- On CIFAR-100-C with ViT-B/16, Mint achieves 44.1% average versus 41.9% for WATT-S (CLIP 35.8%).
- On ImageNet-C with ViT-L/14, Mint achieves 47.0% average versus 43.9% for WATT-S (CLIP 39.6%); Mint runs in 1m07s vs 50m20s for WATT-S on CIFAR-100-C.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply Mint test-time adaptation | ViT-B/32 · CIFAR-10-C | accuracy (Avg., %): +3.9% | Helped |
| apply Mint test-time adaptation | ViT-B/16 · CIFAR-100-C | accuracy (Avg., %): +2.2% | Helped |
| apply Mint test-time adaptation | ViT-L/14 · ImageNet-C | accuracy (Avg., %): +3.1% | Helped |
| reduce Mint batch size to 1 | Mint with ViT-B/16 · CIFAR-100-C | accuracy (%): 35.8 -> \meansd{43.1}{0.1} | Helped |
| remove the gradient accumulator (ablation) | Mint on ViT-B/32 · CIFAR-10-C | accuracy | Hurt |
| add the gradient accumulator (ablation) | Mint on ViT-B/32 · CIFAR-10-C | accuracy | Helped |
| remove the mean accumulator (ablation) | Mint on ViT-B/32 · CIFAR-10-C | accuracy | Hurt |
| add the mean accumulator (ablation) | Mint on ViT-B/32 · CIFAR-10-C | accuracy | Helped |
| remove both accumulators (ablation) | Mint on ViT-B/32 · CIFAR-10-C | accuracy | Hurt |
| update all LayerNorm layers rather than a subset (ablation) | Mint · corruption benchmarks | accuracy | Helped |
| vary batch size from 1 to 200 | Mint · CIFAR-10-C, CIFAR-100-C, and ImageNet-C | accuracy | Mixed |
Surprising findings
- Batch size 1 still yields large gains (e.g., 70.5% on CIFAR-10-C vs 59.0% CLIP), enabled by the mean accumulator.
Limitations
- Analysis focuses on natural corruption and classification tasks; adversarial shifts and dense prediction are not studied.
- Theory is for a binary, simplified RMSNorm-based model with Rademacher latent assumptions and infinite-sample limits.
- Hyperparameters (learning rate, K_prior) are tuned per architecture family, with prior size fixed at 10,000.
Experimental setup
Datasets: CIFAR-10-C CIFAR-100-C ImageNet-C
Benchmarks: CIFAR-10-C CIFAR-100-C ImageNet-C
Metrics: accuracy PL-inter variance GT-inter variance testing time
Baselines: CLIP Ensemble TPT TDA DMN-ZS VTE Zero WATT-S TPS CLIPArTT
Builds on: CLIP TPT WATT CLIPArTT Fisher score
Closest papers by meaning
- Bayesian Test-Time Adaptation for Vision-Language Models (CVPR 2025 · similarity 0.88)
- Controlling Vision-Language Models for Multi-Task Image Restoration (ICLR 2024 · similarity 0.88)
- CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation (NeurIPS 2025 · similarity 0.88)
- Spurious Feature Eraser: Stabilizing Test-Time Adaptation for Vision-Language Foundation Model (AAAI 2025 · similarity 0.87)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Revisiting Visual Corruptions in LVLMs: A Shape-Texture Perspective on Model Failures (CVPR 2026 · similarity 0.87)
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.