Direct Preference Optimization (DPO) has become a standard approach for aligning large language models with human preferences, yet existing methods treat all preference pairs uniformly during training.We identify two distinct sources of learning difficulty: Input Complexity (IC), capturing prompt understanding challenges, and Output Ambiguity (OA), measuring preference discrimination difficulty.Through systematic analysis, we demonstrate that these dimensions induce asymmetric learning dynamics, with IC-related competencies developing rapidly in early training while OA-related competencies emerge more gradually.Building on this observation, we propose DECOPO, a training framework that maintains separate, adaptive pacing schedules for each dimension.Experiments on UltraFeedback show that DE-COPO achieves 42.3% length-controlled win rate on AlpacaEval 2.0 and 7.66 on MT-Bench, outperforming curriculum baselines by 2.1% and 0.21 points respectively, while matching full-data baseline performance with only 75% of training samples.

Label flipping attacks stand as a potent and practical threat to the integrity of machine learning models. While extensive research has focused on designing sophisticated attack and defense mechanisms, the underlying factors that govern a model's susceptibility remain underexplored. This paper reveals a critical phenomenon: the impact of label flipping attacks is highly differential across classes, strongly correlated with the intrinsic confusability between the source and target classes. We provide a rigorous theoretical analysis, demonstrating that a lower standardized separation between classes fundamentally leads to greater vulnerability. Grounded in this insight, we propose Confusability-Aware Contrastive Learning (CACL), a targeted defense that maximizes the feature-space separation for the most vulnerable class pairs. Extensive experiments validate the strong link between class separability and vulnerability, and show that CACL significantly mitigates the attack's impact while providing superior protection for the most susceptible classes. Our code is available at https://github.com/Pinlong-Zhao/Unequal-Vulnerability.

Text sentiment analysis is an important natural language processing (NLP) task and has received considerable attention in recent years. Numerous deep-learning based methods have been proposed in previous literature in terms of new deep neural networks (DNN) including new embedding strategies, new attention mechanisms, and new encoding layers. In this study, an alternative technical path is investigated to further improve the state-of-the-art performance of text sentiment analysis. An new effective learning framework is proposed that combines knowledge distillation and sample selection. A dually-born-again network (DBAN) is presented in which the teacher network and the student network are simultaneously trained through an iterative approach. A selection gate is defined to deal with training samples which are useless or even harmful for model training. Moreover, both the DBAN and sample selection are further improved by ensemble. The proposed framework can improve the existing state-of-the-art DNN models in sentiment analysis. Experimental results indicate that the proposed framework enhances the performances of existing networks. In addition, DBAN outperforms existing born-again network.