Federated Semi-Supervised Learning (FSSL) aims to leverage unlabeled data across clients with limited labeled data to train a global model with strong generalization ability. Most FSSL methods rely on consistency regularization with pseudo-labels, converting predictions from local or global models into hard pseudo-labels as supervisory signals. However, we discover that the quality of pseudo-label is largely deteriorated by data heterogeneity, an intrinsic facet of federated learning. In this paper, we study the problem of FSSL in-depth and show that (1) heterogeneity exacerbates pseudo-label mismatches, further degrading model performance and convergence, and (2) local and global models' predictive tendencies diverge as heterogeneity increases. Motivated by these findings, we propose a simple and effective method called Semi-supervised Aggregation for Globally-Enhanced Ensemble (SAGE), that can flexibly correct pseudo-labels based on confidence discrepancies. This strategy effectively mitigates performance degradation caused by incorrect pseudo-labels and enhances consensus between local and global models. Experimental results demonstrate that SAGE outperforms existing FSSL methods in both performance and convergence. Our code is available at https://github.com/Jay-Codeman/SAGE

In real-world datasets, the challenges of long-tailed distributions and noisy labels often coexist, posing obstacles to the model training and performance. Existing studies on long-tailed noisy label learning (LTNLL) typically assume that the generation of noisy labels is independent of the long-tailed distribution, which may not be true from a practical perspective. In real-world situaiton, we observe that the tail class samples are more likely to be mislabeled as head, exacerbating the original degree of imbalance. We call this phenomenon as ``tail-to-head (T2H)'' noise. T2H noise severely degrades model performance by polluting the head classes and forcing the model to learn the tail samples as head. To address this challenge, we investigate the dynamic misleading process of the nosiy labels and propose a novel method called Disentangling and Unlearning for Long-tailed and Label-noisy data (DULL). It first employs the Inner-Feature Disentangling (IFD) to disentangle feature internally. Based on this, the Inner-Feature Partial Unlearning (IFPU) is then applied to weaken and unlearn incorrect feature regions correlated to wrong classes. This method prevents the model from being misled by noisy labels, enhancing the model's robustness against noise. To provide a controlled experimental environment, we further propose a new noise addition algorithm to simulate T2H noise. Extensive experiments on both simulated and real-world datasets demonstrate the effectiveness of our proposed method.

Data heterogeneity, stemming from local non-IID data and global long-tailed distributions, is a major challenge in federated learning (FL), leading to significant performance gaps compared to centralized learning. Previous research found that poor representations and biased classifiers are the main problems and proposed neural-collapse-inspired synthetic simplex ETF to help representations be closer to neural collapse optima. However, we find that the neural-collapse-inspired methods are not strong enough to reach neural collapse and still have huge gaps to centralized training. In this paper, we rethink this issue from a self-bootstrap perspective and propose FedYoYo (You Are Your Own Best Teacher), introducing Augmented Self-bootstrap Distillation (ASD) to improve representation learning by distilling knowledge between weakly and strongly augmented local samples, without needing extra datasets or models. We further introduce Distribution-aware Logit Adjustment (DLA) to balance the self-bootstrap process and correct biased feature representations. FedYoYo nearly eliminates the performance gap, achieving centralized-level performance even under mixed heterogeneity. It enhances local representation learning, reducing model drift and improving convergence, with feature prototypes closer to neural collapse optimality. Extensive experiments show FedYoYo achieves state-of-the-art results, even surpassing centralized logit adjustment methods by 5.4\% under global long-tailed settings.

Visual prompt tuning (VPT) provides an efficient and effective solution for adapting pre-trained models to various downstream tasks by incorporating learnable prompts. However, most prior art indiscriminately applies a fixed prompt distribution across different tasks, neglecting the importance of each block differing depending on the task. In this paper, we investigate adaptive distribution optimization (ADO) by addressing two key questions: (1) How to appropriately and formally define ADO, and (2) How to design an adaptive distribution strategy guided by this definition? Through in-depth analysis, we provide an affirmative answer that properly adjusting the distribution significantly improves VPT performance, and further uncover a key insight that a nested relationship exists between ADO and VPT. Based on these findings, we propose a new VPT framework, termed PRO-VPT (iterative Prompt RelOcation-based VPT), which adaptively adjusts the distribution building upon a nested optimization formulation. Specifically, we develop a prompt relocation strategy for ADO derived from this formulation, comprising two optimization steps: identifying and pruning idle prompts, followed by determining the optimal blocks for their relocation. By iteratively performing prompt relocation and VPT, our proposal adaptively learns the optimal prompt distribution, thereby unlocking the full potential of VPT. Extensive experiments demonstrate that our proposal significantly outperforms state-of-the-art VPT methods, e.g., PRO-VPT surpasses VPT by 1.6% average accuracy, leading prompt-based methods to state-of-the-art performance on the VTAB-1k benchmark. The code is available at https://github.com/ckshang/PRO-VPT.

Effectively handling the co-occurrence of non-IID data and long-tailed distributions remains a critical challenge in federated learning. While fine-tuning vision-language models (VLMs) like CLIP has shown to be promising in addressing non-IID data challenges, this approach leads to severe degradation of tail classes in federated long-tailed scenarios. Under the composite effects of strong non-IID data distribution and long-tailed class imbalances, VLM fine-tuning may even fail to yield any improvement. To address this issue, we propose Class-Aware Prompt Learning for Federated Long-tailed Learning (CAPT), a novel framework that leverages a pre-trained VLM to effectively handle both data heterogeneity and long-tailed distributions. CAPT introduces a dual-prompt mechanism that synergizes general and class-aware prompts, enabling the framework to capture global trends while preserving class-specific knowledge. To better aggregate and share knowledge across clients, we introduce a heterogeneity-aware client clustering strategy that groups clients based on their data distributions, enabling efficient collaboration and knowledge sharing. Extensive experiments on various long-tailed datasets with different levels of data heterogeneity demonstrate that CAPT significantly improves tail class performance without compromising overall accuracy, outperforming state-of-the-art methods in federated long-tailed learning scenarios.

In this report, we present the third technical report on the development of slow-thinking models as part of the STILL project. As the technical pathway becomes clearer, scaling RL training has become a central technique for implementing such reasoning models. We systematically experiment with and document the effects of various factors influencing RL training, conducting experiments on both base models and fine-tuned models. Specifically, we demonstrate that our RL training approach consistently improves the Qwen2.5-32B base models, enhancing both response length and test accuracy. Furthermore, we show that even when a model like DeepSeek-R1-Distill-Qwen-1.5B has already achieved a high performance level, it can be further refined through RL training, reaching an accuracy of 39.33% on AIME 2024. Beyond RL training, we also explore the use of tool manipulation, finding that it significantly boosts the reasoning performance of large reasoning models. This approach achieves a remarkable accuracy of 86.67% with greedy search on AIME 2024, underscoring its effectiveness in enhancing model capabilities. We release our resources at the STILL project website: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs.

Automatic X-ray prohibited item detection is vital for public safety. Existing deep learning-based methods all assume that the annotations of training X-ray images are correct. However, obtaining correct annotations is extremely hard if not impossible for large-scale X-ray images, where item overlapping is ubiquitous.As a result, X-ray images are easily contaminated with noisy annotations, leading to performance deterioration of existing methods.In this paper, we address the challenging problem of training a robust prohibited item detector under noisy annotations (including both category noise and bounding box noise) from a novel perspective of data augmentation, and propose an effective label-aware mixed patch paste augmentation method (Mix-Paste). Specifically, for each item patch, we mix several item patches with the same category label from different images and replace the original patch in the image with the mixed patch. In this way, the probability of containing the correct prohibited item within the generated image is increased. Meanwhile, the mixing process mimics item overlapping, enabling the model to learn the characteristics of X-ray images. Moreover, we design an item-based large-loss suppression (LLS) strategy to suppress the large losses corresponding to potentially positive predictions of additional items due to the mixing operation. We show the superiority of our method on X-ray datasets under noisy annotations. In addition, we evaluate our method on the noisy MS-COCO dataset to showcase its generalization ability. These results clearly indicate the great potential of data augmentation to handle noise annotations. The source code is released at https://github.com/wscds/Mix-Paste.

Mental health risk prediction is a growing field in the speech community, but many studies are based on small corpora. This study illustrates how variations in test and train set sizes impact performance in a controlled study. Using a corpus of over 65K labeled data points, results from a fully crossed design of different train/test size combinations are provided. Two model types are included: one based on language and the other on speech acoustics. Both use methods current in this domain. An age-mismatched test set was also included. Results show that (1) test sizes below 1K samples gave noisy results, even for larger training set sizes; (2) training set sizes of at least 2K were needed for stable results; (3) NLP and acoustic models behaved similarly with train/test size variations, and (4) the mismatched test set showed the same patterns as the matched test set. Additional factors are discussed, including label priors, model strength and pre-training, unique speakers, and data lengths. While no single study can specify exact size requirements, results demonstrate the need for appropriately sized train and test sets for future studies of mental health risk prediction from speech and language.

Machine learning models for speech-based depression classification offer promise for health care applications. Despite growing work on depression classification, little is understood about how the length of speech-input impacts model performance. We analyze results for speaker-independent depression classification using a corpus of over 1400 hours of speech from a human-machine health screening application. We examine performance as a function of response input length for two NLP systems that differ in overall performance. Results for both systems show that performance depends on natural length, elapsed length, and ordering of the response within a session. Systems share a minimum length threshold, but differ in a response saturation threshold, with the latter higher for the better system. At saturation it is better to pose a new question to the speaker, than to continue the current response. These and additional reported results suggest how applications can be better designed to both elicit and process optimal input lengths for depression classification.

Digital screening and monitoring applications can aid providers in the management of behavioral health conditions. We explore deep language models for detecting depression, anxiety, and their co-occurrence from conversational speech collected during 16k user interactions with an application. Labels come from PHQ-8 and GAD-7 results also collected by the application. We find that results for binary classification range from 0.86 to 0.79 AUC, depending on condition and co-occurrence. Best performance is achieved when a user has either both or neither condition, and we show that this result is not attributable to data skew. Finally, we find evidence suggesting that underlying word sequence cues may be more salient for depression than for anxiety.