# Rethinking Noise-Robust Training for Frozen Vision Foundation Models: A Cross-Dataset Benchmark with a Case Study of Small-Loss Failure

arXiv: https://arxiv.org/abs/2605.22591 · HTML: https://mlsanity.com/paper/2605.22591/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 11 of its 16 references that are in the map.
- First posted on arXiv 21 May 2026.

## Builds on (most cited first, 11 of 11)

- [DINOv2: Learning Robust Visual Features without Supervision](https://mlsanity.com/paper/2304.07193/) (2023)
- [Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels](https://mlsanity.com/paper/1804.06872/) (2018)
- [When Does Label Smoothing Help?](https://mlsanity.com/paper/1906.02629/) (2019)
- [A Closer Look at Memorization in Deep Networks](https://mlsanity.com/paper/1706.05394/) (2017)
- [MedMNIST v2 -- A large-scale lightweight benchmark for 2D and 3D biomedical image classification](https://mlsanity.com/paper/2110.14795/) (2021)
- [Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)](https://mlsanity.com/paper/1902.03368/) (2019)
- [BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs](https://mlsanity.com/paper/2303.00915/) (2023)
- [Early-Learning Regularization Prevents Memorization of Noisy Labels](https://mlsanity.com/paper/2007.00151/) (2020)
- [Towards Understanding Deep Learning from Noisy Labels with Small-Loss Criterion](https://mlsanity.com/paper/2106.09291/) (2021)
- [When Noisy Labels Meet Long Tail Dilemmas: A Representation Calibration Method](https://mlsanity.com/paper/2211.10955/) (2022)
- [Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise](https://mlsanity.com/paper/2412.00150/) (2024)

## Abstract (arXiv)

Frozen Vision Foundation Models (VFMs) with lightweight classification heads are increasingly used in medical imaging because they offer efficient and reproducible deployment. Yet noisy-label learning methods for this frozen-feature regime remain poorly understood, and most existing methods still rely on a small-loss assumption inherited from end-to-end training. We present a controlled benchmark of eight noisy-label methods across five medical datasets, three backbones, two noise types, and five noise rates (150 conditions, 6,000 training runs), evaluated with balanced accuracy. The benchmark shows that there is no universal winner: Friedman ranking over the 150 conditions yields chi2 = 333.2 (p = 4.77 × 10-68), ELR wins the most conditions (49/150), while CUFIT attains the best mean rank (2.51). The practical cost of method choice grows sharply with noise severity, from 4.5pp on clean data to 18.8pp at asymmetric 40% noise. To explain these benchmark-level patterns, we revisit the small-loss assumption in a representative high-risk regime. Under frozen DINOv2 features, clean and noisy loss distributions overlap by 53--61%, and matched-rate clean-sample detection shows that prediction agreement is markedly more stable than loss ranking under asymmetric noise (3pp vs. 13pp precision drop). On ISIC2019 with asymmetric 40% noise, Co-Teaching reaches 68% overall accuracy while collapsing to 35.1% balanced accuracy with zero recall on three minority classes. Together, these results recast noisy-label learning for frozen VFMs as a regime-aware method-selection problem rather than a search for a single dominant algorithm. We conclude with evidence-based guidance and a low-regret feature-space selector for practical recommendation.