Parameter-Efficient Fine-Tuning (PEFT) has emerged as the dominant paradigm for adapting Vision-Language Models (VLMs), yet its robustness to label noise remains largely unexplored. To address this gap, we conduct a systematic evaluation of representative PEFT methods across eight datasets under diverse noise conditions. Our experiments reveal a consistent robustness hierarchy: methods like Tip-Adapter-F and LoRA demonstrate high resilience, whereas Linear Probing suffers significant degradation. We attribute this disparity to a mechanism we term Gradient-Loss Decoupling. Our analysis shows that robust architectures maintain a “loose coupling" between prediction error and gradient magnitude, effectively preventing high-loss outliers from generating disruptive parameter updates. We substantiate this through mathematical derivation, proving that specific architectural constraints, such as affinity-based gating, naturally suppress noise gradients. Finally, we empirically validate this framework through targeted interventions that enforce decoupling to recover robustness, confirming that gradient dynamics, rather than simple factors like model capacity, are the key drivers of noise resilience in VLM adaptation.

We propose a novel sample selection method for image classification in the presence of noisy labels. Existing methods typically consider small-loss samples as correctly labeled. However, some correctly labeled samples are inherently difficult for the model to learn and can exhibit high loss similar to mislabeled samples in the early stages of training. Consequently, setting a threshold on per-sample loss to select correct labels results in a trade-off between precision and recall in sample selection: a lower threshold may miss many correctly labeled hard-to-learn samples (low recall), while a higher threshold may include many mislabeled samples (low precision). To address this issue, our goal is to accurately distinguish correctly labeled yet hard-to-learn samples from mislabeled ones, thus alleviating the trade-off dilemma. We achieve this by considering the trends in model prediction confidence rather than relying solely on loss values. Empirical observations show that only for correctly labeled samples, the model's prediction confidence for the annotated labels typically increases faster than for any other classes. Based on this insight, we propose tracking the confidence gaps between the annotated labels and other classes during training and evaluating their trends using the Mann-Kendall Test. A sample is considered potentially correctly labeled if all its confidence gaps tend to increase. Our method functions as a plug-and-play component that can be seamlessly integrated into existing sample selection techniques. Experiments on several standard benchmarks and real-world datasets demonstrate that our method enhances the performance of existing methods for learning with noisy labels.

Fine-grained entity typing (FET) aims to assign proper semantic types to entity mentions according to their context, which is a fundamental task in various entity-leveraging applications. Current FET systems usually establish on large-scale weaklysupervised/distantly annotation data, which may contain abundant noise and thus severely hinder the performance of the FET task. Although previous studies have made great success in automatically identifying the noisy labels in FET, they usually rely on some auxiliary resources which may be unavailable in real-world applications (e.g., pre-defined hierarchical type structures, humanannotated subsets). In this paper, we propose a novel approach to automatically correct noisy labels for FET without external resources. Specifically, it first identifies the potentially noisy labels by estimating the posterior probability of a label being positive or negative according to the logits output by the model, and then relabel candidate noisy labels by training a robust model over the remaining clean labels. Experiments on two popular benchmarks prove the effectiveness of our method. Our source code can be obtained from https://github.com/CCIIPLab/DenoiseFET.