data augmentation

To address the challenges in domain generalization—namely, unknown test distributions and semantic degradation caused by aggressive data augmentation—this paper proposes a robust adaptive augmentation framework. Methodologically, it (1) uniformly samples diverse strong augmentations (e.g., Cutout, ColorJitter) to eliminate heuristic bias in intensity design; (2) introduces, for the first time, an unsupervised, learnable gradient-aware rejection mechanism that dynamically filters out harmful augmented samples without target-domain labels; and (3) designs a reinforcement learning–inspired reward function to jointly optimize augmentation strength and semantic fidelity. Evaluated on standard benchmarks including PACS and Office-Home, the method achieves state-of-the-art (SOTA) or SOTA-comparable performance, with average accuracy improvements of 1.8–3.2% and significantly enhanced generalization robustness.

To address the weak generalization and severe prediction bias of large language models (LLMs) on few-shot and class-imbalanced text data, this paper proposes an embedding-space synthetic feature augmentation method. Unlike conventional approaches, it operates directly in the language model’s latent embedding space—bypassing raw text generation—and jointly synthesizes minority-class features via embedding interpolation, noise perturbation, and adversarial generation to optimize semantic representation distributions. The method integrates seamlessly into standard fine-tuning pipelines and is compatible with mainstream open-source text classification benchmarks. Experiments across multiple benchmarks demonstrate up to a 12.3% improvement in minority-class F1 score, alongside consistent gains in overall accuracy and robustness. The core innovation lies in migrating synthetic data generation from the input space to the embedding space, enabling efficient, lossless, and fair representation calibration.

This work addresses test-time adaptation (TTA), aiming to enable models to adapt in real time to diverse, unseen target-domain distributions without fine-tuning or online parameter updates, while preserving source-domain knowledge. To this end, we propose GeneralizeFormer—a lightweight meta-learning Transformer architecture that dynamically generates only BatchNorm layer parameters and classifier weights, thereby improving computational efficiency and enhancing source-feature retention. We further introduce a layer-wise gradient-aware mechanism to strengthen robustness against distribution shifts. Evaluated on six mainstream domain generalization benchmarks, GeneralizeFormer significantly outperforms existing TTA and domain generalization methods. It demonstrates strong efficacy in handling multiple heterogeneous target domains, dynamic environments, and continual distribution drift, establishing new state-of-the-art performance under realistic test-time adaptation settings.

Distribution shift severely degrades the cross-domain performance of deep models, while existing domain generalization (DG) methods often rely on multi-source domain training or computationally expensive test-time adaptation. To address this, we propose a lightweight, training-free test-time augmentation method: during inference, input images are transformed via elastic deformation and grid warping to generate class-invariant variants; high-confidence predictions are then dynamically selected based on model confidence scores for ensemble aggregation. Our approach introduces only minimal deformable augmentations and a confidence-guided fusion mechanism—requiring no additional training, and offering plug-and-play compatibility with diverse DG algorithms and backbone architectures. Evaluated on PACS and Office-Home benchmarks, it consistently improves the cross-domain generalization of multiple state-of-the-art models, yielding average accuracy gains of 2.1–4.7%, and significantly enhancing robustness to unseen domains.

Synthetic image data often exhibits inconsistent quality and vastly exceeds real-data quantities, leading to domain shift and degraded model performance. Method: We propose a fine-tuning-free synthetic data augmentation framework for diffusion models. (1) Leveraging joint disambiguation of class names by LLMs and CLIP, we design contextualized and stylized prompting strategies to enhance semantic fidelity and diversity of generated images; (2) an auxiliary batch normalization module is introduced to mitigate distributional shift between synthetic and real domains. Contribution/Results: We provide the first empirical evidence that off-the-shelf generative models—without adaptation—can substantially boost recognition accuracy. Our diversity-aware generation strategy sustains consistent performance gains even when scaling ImageNet to six times its original size. The method achieves significant classification accuracy improvements on both ImageNet and cross-domain benchmarks, demonstrating strong generalization capability.

To address the poor robustness, high computational cost, and pretraining knowledge degradation (e.g., due to Dropout) of fine-tuned models under distribution shifts, this paper proposes High-Ratio Mixout: a fine-tuning strategy that replaces up to 90% of learnable parameters with their corresponding pretrained weights—enforcing strong regularization while enabling adaptive parameter updates. Applied dynamically to ViT and ResNet architectures, it balances pretrained knowledge retention and domain generalization. Evaluated on five standard out-of-distribution benchmarks (PACS, VLCS, etc.), our method achieves single-model performance competitive with ensemble approaches, reduces gradient computation by 45%, and cuts GPU memory usage by 90%, significantly improving training efficiency. Its core contribution lies in the first systematic investigation of high-masking-rate Mixout for robust fine-tuning, revealing its mechanism for enhancing generalization. The result is a lightweight, efficient, and plug-and-play paradigm for out-of-distribution generalization.

This study investigates the mechanism by which synthetic data augmentation improves score-based classification performance—measured by metrics such as AUROC and AUPRC—in class-imbalanced settings. By developing a theoretical framework that disentangles the effects of augmentation on effective class weighting and distributional bias, and integrating tools from statistical learning theory, minimax analysis, and finite-sample error decomposition, the work establishes that under correctly specified models, augmentation solely reduces variance without improving overall performance. However, under model misspecification, it can mitigate ranking errors by correcting class imbalance. The analysis yields novel minimax lower bounds, which are corroborated through simulation experiments.

To address the dual challenges of intra-domain class imbalance and cross-domain distribution shift in Domain-Incremental Learning (DIL), this paper proposes the Dual-Balanced Collaborative Experts (DBCE) framework. DBCE employs frequency-aware expert grouping and dynamic expert selection to enable specialized modeling of classes with varying frequencies. It further introduces a balanced Gaussian sampling strategy for pseudo-feature generation, leveraging historical statistics to mitigate underlearning of minority classes and cross-domain knowledge degradation. A dedicated loss function is designed to jointly optimize knowledge retention and class balance. Evaluated on four benchmark datasets, DBCE significantly improves minority-class accuracy while effectively suppressing catastrophic forgetting. It achieves dynamic performance equilibrium between majority and minority classes across both old and new domains, thereby enhancing the robustness and adaptability of DIL in realistic imbalanced scenarios.

This work addresses the limited generalization of prompt-based large language model (LLM) classifiers in data-scarce scenarios, where insufficient fine-tuning often hinders performance. The authors propose a multi-task prompt fine-tuning approach that designs task-specific prompts while integrating general instruction tuning, substantially improving classification accuracy on unseen domains and novel prompts. Notably, they find that supervised classification training without explicit reasoning capabilities can effectively generalize to reasoning-intensive tasks such as summarization. To mitigate performance degradation caused by prompt variations, a hybrid training strategy is introduced. Experimental results demonstrate strong performance on related unseen tasks, highlighting the potential of classification-oriented training for building versatile, general-purpose monitoring systems.

This work addresses the lack of rigorous theoretical foundations for domain alignment mechanisms in semi-supervised domain adaptation. We propose the first unified generalization error analysis framework, characterizing the sample complexity of deep domain adaptation networks based on both Maximum Mean Discrepancy (MMD) and adversarial objectives. Methodologically, we jointly model feature transformation and classifier learning, quantify function class complexity via covering number theory, and integrate MMD and adversarial losses to derive a data-dependent generalization upper bound. Theoretically, we establish that sample complexity grows quadratically with network depth and width, and that target-domain risk scales with the inverse square root of labeled target samples—thereby enhancing robustness under limited labeling. Empirical evaluations validate these theoretical predictions, providing an interpretable, principled foundation for modern deep domain adaptation methods.