Real-world data typically exhibits long-tailed class distribution and contains label noise. Previous long-tail learning methods overlooked the prevalence of noisy labels in training. Moreover, the commonly used small-loss noisy label detection criterion fails in long-tail data.

To solve the problems, a research team led by Professor Yu-Feng Li and Professor Min-Ling Zhang published their new research on 15 January 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed a novel framework called RoLT, which contains a novel class-wise small-distance criterion to detect noisy labels and select clean training samples. It also incorporates soft pseudo-labeling with label distribution to enhance training for tail classes. Compared with the existing research results, the proposed framework achieves higher model performance under different settings. The improvement becomes more significant at high noise levels.

In the research, they first study the commonly used small-loss noisy label detection criterion. Due to the misclassification of samples with tail classes (large losses), the small-loss criterion cannot distinguish clean samples of tail classes and samples with noisy labels. Contrastively, the class-wise small-distance criterion can be more robust. Specifically, for each individual class, it selects small-distance samples as clean where distance is calculated between the sample and its class prototype in the embedding space. The empirical study demonstrates that the class-wise small-distance criterion is more robust than the small-loss criterion, and works well on both head and tail classes.

Motivated by this, the researchers propose a robust framework called RoLT, which incorporates soft pseudo-labeling with label distribution to enhance training for tail classes. Extensive experiments on multiple benchmark datasets clearly demonstrate that RoLT outperforms previous methods under different settings. The improvement becomes more significant at high noise levels.

The main contributions of this paper can be summarized as follows: (1) This paper addresses the underexplored problem of learning from long-tail data in the presence of label noise, taking a significant step toward real-world applications. (2) This paper identifies the limitations of the widely used small-loss criterion and introduce a novel criterion called class-wise small-distance. Building upon this, the authors propose a robust framework called RoLT. (3) Extensive experiments conducted on benchmark and real-world datasets demonstrate the superiority of our proposed method.

The long-tailed distribution and label noise widely exist in real-world data. Future work can exploit more public datasets, expanding the learning framework towards more research fields, and designing more efficient, accurate, and scalable method for robust long-tail learning.