Just Y-Prediction: Enabling Historical Cumulative Inconsistency in Label Diffusion for Learning with Noisy Label

Label noise is pervasive in real-world datasets and significantly compromises model generalization, fueling extensive research into Learning with Noisy Labels (LNL). Most LNL methods focus on robust discriminative learning, while recent generative classifiers such as label diffusion models (LDMs) show superior robustness by modeling class posteriors. However, current LDMs predominantly rely on standard $\epsilon$-prediction, where Gaussian pnoise lacks explicit class semantics, limiting both optimization and inference under label noise environments. To address this issue, we propose just y-prediction (JYP), a novel training paradigm that enables LDMs to directly characterize the label manifold and leverage explicit class-semantic guidance. Theoretically, we prove that JYP converges to an optimal solution equivalent to that of $\epsilon$-prediction within the label diffusion framework, while facilitating accelerated convergence and enabling one-step inference. Leveraging JYP as a foundation, we further incorporate historical cumulative inconsistency to adaptively tailor optimization strategies for clean, noisy, and hard samples. Extensive experiments demonstrate that our method consistently outperforms competitors across diverse synthetic noisy datasets and achieves state-of-the-art performance on multiple real-world benchmarks.

Machine learning systems usually learn from labeled examples, but in many real-world datasets some labels are wrong. These mistakes can come from crowdsourcing, web scraping, or automatic labeling tools, and they may cause models to learn incorrect patterns. Our work aims to make classification models more reliable when the training labels are noisy. We introduce Just Y-Prediction (JYP), a method that uses a generative label diffusion model to directly recover clean labels from corrupted label information. Unlike standard diffusion methods that predict random noise step by step, JYP predicts the label itself, which gives the model clearer guidance for classification. We also track how consistently each sample disagrees with its given label during training, helping the model distinguish likely clean samples, likely noisy samples, and difficult samples. Across a wide range of noisy-label benchmarks, JYP achieves strong performance and supports efficient one-step prediction. Our results show that directly modeling labels can be a practical way to improve learning from imperfect data.