Tong Wei, Jiang-Xin Shi, Min-Ling Zhang, Yu-Feng Li
原文摘要
Long-tailed learning aims to enhance the generalization performance of underrepresented tail classes. However, previous methods have largely overlooked the prevalence of noisy labels in training data. In this paper, we address the challenge of noisy labels in long-tailed learning. We identify a critical issue: the commonly used small-loss noisy label detection criterion fails to perform effectively in long-tailed class distributions. This failure arises from the inherent bias of deep neural networks, which tend to misclassify tail class examples as head classes, leading to unreliable loss calculations. To mitigate this, we propose a novel small-distance criterion that leverages the robustness of learned representations, enabling more accurate identification of correctly-labeled examples across both head and tail classes. Additionally, to improve training for tail classes, we replace discrete pseudo-labels with label distributions for examples flagged as noisy, resulting in significant performance gains. Based on these contributions, we introduce the Robust Long-Tail learning framework, designed to train models that are resilient to both class imbalance and noisy labels. Extensive experiments on benchmark and real-world datasets demonstrate that our approach outperforms previous methods, offering substantial performance improvements.
Mao-Lin Luo, Zhe-Xu Wang, Zi-Hao Zhou, Bo Ye, Jian Zhao, Min-Ling Zhang, Tong Wei
原文摘要
Continual post-training is becoming a central paradigm for adapting vision-language models to evolving tasks. Recent studies report that reinforcement learning is less prone to forgetting than supervised fine-tuning, motivating the view that RL is inherently resistant to forgetting. However, this view remains insufficiently validated, as existing evidence is largely drawn from outdated or homogeneous benchmarks. We revisit this assumption by introducing MRCL, a Multimodal Reasoning Continual Learning benchmark built from recent and diverse multimodal reasoning tasks. Experiments on MRCL show that standard reinforcement learning still suffers from catastrophic forgetting during continual post-training. We trace the failure to an objective mismatch. The KL regularization used in common policy optimization methods is evaluated on current-task data, whereas forgetting is caused by behavioral drift on prior-task distributions. To address this problem, we propose Continual Policy Optimization (CPO), a replay-free method grounded in a prior-task behavioral KL objective. CPO derives a local Fisher surrogate from the historical KL objective and uses parameter movement as a gradient-free proxy for Fisher sensitivity, enabling sparse regularization with negligible additional overhead. Experiments on three model scales and comparisons with multiple RL baselines show that CPO consistently reduces forgetting while maintaining effective adaptation and preserving broader pretrained capabilities. The implementation code is available at https://github.com/MaolinLuo/CPO.