Meteorological forecasting under extreme uncertainty remains challenging because purely data-driven models often produce unstable and physically implausible predictions in rare and severe weather events. Although physics-informed machine learning (PIML) has been introduced to alleviate this issue, most existing methods incorporate physical constraints directly into model architectures or training objectives, limiting flexibility across heterogeneous forecasting models. To address this problem, we propose the Adaptive Post-hoc Knowledge Consistency (APKC) framework, a lightweight and model-agnostic method that imposes temporal, spatial, and physical consistency directly on model outputs without modifying backbone structures. In addition, APKC employs an adaptive weighting mechanism to adjust constraint strength under different uncertainty levels. Experiments on the ERA5 reanalysis dataset, including extreme-event subsets, show that APKC consistently improves prediction stability, numerical accuracy, and physical plausibility across diverse time-series forecasting models, with negligible or even positive impact under normal conditions.

Federated learning (FL) enables multiple clients to collaboratively build a shared model without exchanging raw data, and has become an important paradigm for privacy-preserving distributed learning. However, in scenarios with significant data distribution imbalance among clients, severe client heterogeneity often induces local distribution shift and unstable optimization, resulting in degraded model performance. To address this issue, this paper proposes FedHAM, a heterogeneity-aware collaborative optimization method for federated learning. Different from treating data augmentation and knowledge regularization as isolated components, FedHAM leverages client heterogeneity as a unified driving signal to jointly regulate data-level expansion and output-level regularization. Specifically, a heterogeneity estimation module is designed to characterize the discrepancy between local and global distributions, based on which the Mixup intensity is adaptively adjusted to alleviate class bias and sample sparsity. Furthermore, a global-model-informed distillation framework is adopted to mitigate unstable predictions under data heterogeneity and to stabilize the local training process. Extensive experimental analysis under different heterogeneous scenarios indicates that the proposed framework achieves more stable optimization and stronger generalization ability compared with representative federated learning approaches.

This paper proposes a robust Federated Learning (FL) framework, FedVC-ADDiM for the efficient diagnosis of Alzheimer’s disease using decentralized MRI brain scans. The framework integrates a multi-objective optimization technique, Pareto Front Selection for Multi-Objective Optimization (PFSMOO), to select the most suitable global-local model from a diverse set of pre-trained architectures, with the VGG19 model identified as the optimal choice. To enhance the model FL capabilities, we introduce an attention-augmented Convolutional Block Attention Module (CBAM) within the VGG19 architecture, enabling the model to focus on critical neuroimaging features. Existing models often suffer from issues such as overfitting due to deep architectures, which can lead to poor generalization on unseen data, and the inability to effectively prioritize efficient features, resulting in reduced accuracy in complex medical diagnoses. The FedVC-ADDiM framework addresses these challenges by leveraging federated knowledge sharing, ensuring data privacy while maintaining model accuracy across diverse, decentralized datasets. Experimental results show that the model with CBAM achieved a notable increase in overall accuracy (95.68

In modern high-density pig farming scenarios, automatic pig detection is severely challenged by frequent occlusions, high visual similarity among individuals, and dynamic environmental interferences, leading to unstable feature extraction and limited model generalization. To address these issues, this study proposes a collaborative detection framework integrating contrastive learning with reinforcement learning. Specifically, a geometry-enhanced contrastive learning module is first constructed, in which affine transformations and random occlusions are applied to generate sample pairs. Based on the triplet loss originally proposed in FaceNet, a margin term is introduced to explicitly enforce intra-class compactness and inter-class separation, thereby enhancing feature robustness against occlusion and deformation. Furthermore, a reinforcement learning decision mechanism based on proximal policy optimization (PPO) is designed. This mechanism incorporates dynamic threshold adjustment and false-positive suppression strategies, and employs a multi-objective reward function to drive the adaptive optimization of detection parameters. Experiments on a dataset of 500 top-view images, using an 8:2 train-validation split, show that the proposed method achieves an mAP50-95 of 0.80, outperforming the YOLOv11-n baseline (0.72). These results demonstrate the effectiveness and robustness of the proposed approach for dense pig detection in comparable farm environments.

Aspect Sentiment Triplet Extraction (ASTE) aims to jointly identify aspect terms, opinion terms, and the sentiment polarity expressed between them in review sentences. Existing tagging-based and table-filling-based methods usually represent token-pair relations with a dense two-dimensional grid. However, only a limited portion of grid cells contributes to valid sentiment triplets, while many irrelevant interactions may introduce noise and weaken robustness, especially in cases involving long-distance associations, multi-word expressions, and overlapping triplets. To address this problem, we propose TASIS, a triplet-aware sparse interaction framework for ASTE. TASIS first uses an adaptive semantic refinement module to assign soft importance weights to token-pair interactions and reduce the influence of uninformative cells. It then applies a semantic-selective Mamba encoder based on state space models to capture structured dependencies over the refined relation grid. Finally, a boundary-driven decoder recovers aspect–opinion spans and predicts their sentiment labels. Experiments on four benchmark datasets show that TASIS achieves consistently strong performance against representative baselines, with clearer advantages in complex relational scenarios. These results indicate that sparse, triplet-aware interaction modeling is effective for structured sentiment extraction.

We address the challenge of semantic-geometric inconsistency in generating 3D models from 2D illustrations, a problem stemming from the lack of explicit structural knowledge in standard generative pipelines. We propose a knowledge-driven framework that integrates machine learning with explicit knowledge representation. Specifically, we introduce a Structural Knowledge Extraction mechanism that decomposes single-view illustrations into semantically distinct layers (foreground/background), effectively converting unstructured image data into structured semantic priors. These priors guide a Knowledge-Guided Diffusion Inference process, utilizing a novel Layer-CLIP module to align visual features with semantic concepts in a shared latent space. This approach facilitates knowledge disentanglement, minimizing the mutual interference of attributes (e.g., color bleeding) and enhancing the fidelity of the reconstructed 3D topology. By fusing explicit layer-wise knowledge with implicit latent diffusion, our method significantly reduces generation hallucinations and improves semantic consistency without relying on massive 3D datasets.

Named entity recognition (NER) aims to distinguish nested entities that share boundary tokens accurately. Existing span refinement methods typically rely on word-pair representation matrices and standard convolutions, but isotropic smoothing blurs boundary semantics and weakens discrimination. To address this, we propose Difference-Aware Boundary Refinement (DABR). DABR enhances boundary distinction by explicitly modeling feature differences between adjacent spans using a hybrid D-Mixer module with local and global paths. The local path employs directional difference convolutions to capture start-boundary shifts and end-boundary transitions. In contrast, the global path introduces a lightweight dual-gated attention mechanism with spatial and channel gating to emphasize salient boundary regions and discriminative semantic channels. Sentence-level context is used to validate candidate boundaries. During inference, structural reparameterization merges multi-path differential computations into a single convolution for efficiency. Extensive experiments on ACE2004, ACE2005, GENIA, CoNLL2003, and OntoNotes 5.0 show that DABR achieves state-of-the-art or highly competitive results across all settings.

Named entity recognition (NER) is a fundamental task in natural language processing. However, traditional methods often face challenges when recognizing nested entities due to blurred boundaries and limited span enumeration. To address these issues, we propose an entity boundary denoising training paradigm that gradually refines entity boundaries through noise-aware span proposals, aiming to improve nested entity recognition. Specifically, we introduce controllable noise into learnable span proposals via dynamic convolution to generate noise-aware span proposals. We then utilize a cross-attention mechanism to dynamically aggregate sentence features, thereby building context-aware representations that better capture global dependencies and enhance semantic alignment between candidate spans and gold entities, which helps mitigate semantic drift caused by noise. Furthermore, we design a selective self-attention mechanism based on inter-query sparse similarity to suppress redundant information, enabling the model to focus on key features. Additionally, we develop a hybrid dynamic supervision method that combines one-to-one and one-to-many label assignments to enhance decoder supervision and improve the quality of candidate proposals. Experimental results on six benchmark NER datasets demonstrate that our approach achieves competitive performance, validating the effectiveness of the proposed method.

Stratospheric balloon station-keeping in highly uncertain wind fields poses a critical challenge due to tail-risk-dominated dynamics, where rare but severe disturbances can lead to catastrophic deviations. Conventional reinforcement learning methods, which optimize expected returns, often fail to ensure robustness under such heavy-tailed conditions. In this paper, we establish a risk-sensitive decision-making framework based on distributional reinforcement learning to explicitly address tail risks in underactuated balloon control. Building upon Implicit Quantile Networks (IQN), we introduce an exponential spectral risk weighting mechanism that enables continuous distortion of return distributions in the quantile space. Unlike conventional CVaR-based approaches relying on hard truncation, the proposed method provides a smooth and flexible risk control paradigm without modifying the underlying learning objective. Extensive experiments in the Balloon Learning Environment demonstrate that the proposed method consistently improves robustness, achieving superior tail performance (CVaR@5

Cryo-electron microscopy (cryo-EM) is a key technique for determining near-atomic resolution structures of biological macromolecules. The reconstructed three-dimensional models not only represent physical conformations but also constitute high-dimensional structured knowledge in life sciences. Existing deep learning methods struggle to extract complete structural knowledge efficiently and robustly from large-scale, high-noise cryo-EM images. To address this, we propose cryoHIRI, a framework that integrates the local feature perception of multi-scale convolutions with the global context modeling of Transformers to achieve coherent understanding of multi-scale image features. Experiments show that cryoHIRI accurately recovers fine-grained structures of complex macromolecules under high noise conditions. It provides an efficient and reliable approach for structural knowledge discovery and supports downstream knowledge engineering tasks.

Accurate short-term temperature forecasting is challenging due to strong non-stationarity and nonlinear atmospheric dynamics. Existing methods often lack strict temporal causality, effective integration of physical priors, and robust multi-scale modeling. To address these issues, we propose ThermoCausalNet, a physics-constrained hybrid framework. It consists of three components: (1) a Causal Dynamic Gated Convolution (CDGC) module for causality-preserving local feature extraction with adaptive kernels; (2) a Meteorology–Physics Dual-Path Adaptive Attention (MPDAttention) module for unified fusion of meteorological data and physical priors; and (3) a Frequency-Enhanced Channel Attention Module (FECAM) for joint temporal–frequency representation learning. Extensive experiments demonstrate that ThermoCausalNet consistently outperforms state-of-the-art baselines across multiple horizons, achieving lower forecasting errors and validating the effectiveness of causality preservation, physics-guided fusion, and multi-scale modeling.

Federated learning enables distributed learning by allowing data to remain on local devices and aggregating the model updates, thereby reducing the risk of data leakage and enhancing privacy protection. However, there are several challenges in federated learning. If participants do not receive appropriate rewards, they may lack the incentive to participate in federated learning tasks. Additionally, the federated learning system may be vulnerable to Byzantine client attackers who intentionally disrupt the model training process and compromise overall performance. Traditional federated learning assumes independent and identically distributed (IID) data, which fails to accommodate data heterogeneity effectively. Therefore, there is an urgent need for an adaptive incentive mechanism explicitly designed for Non-IID data environments. In this paper, we propose FRIFL, a federated learning incentive mechanism that rewards participants based on their reputation and contributions. Considering that each participant may possess data with different distributions and characteristics, we design an attack detection module and account for data heterogeneity. Experimental results demonstrate that FRIFL ensures robustness in unreliable environments with both IID and Non-IID data distributions. Besides, it provides security for the FL system to ensure incentive fairness.

Cryo-electron microscopy (cryo-EM) single-particle reconstruction is challenged by molecular flexibility and unknown imaging poses when resolving heterogeneous conformations. We propose CryoHCT, a Transformer-based encoder-decoder framework for ab initio heterogeneous reconstruction without pose priors. By integrating a conformation Transformer with an implicit neural decoder, CryoHCT achieved mean pose errors of 2.0° (IgG-1D) and 1.9° (Spike-MD), outperformed amortized inference baselines. Fourier Shell Correlation (FSC at 0.5 threshold) resolutions of 11.79 Å and 6.50 Å confirm high-fidelity recovery of continuous dynamics (e.g., Fab rotation). The method bridges pose diversity exploration with conformational continuity, offering a scalable solution for dynamic biomolecule analysis.

Named entity recognition (NER) is a key task in natural language processing, but existing methods often fail to effectively handle nested structures due to fuzzy entity boundaries and structural ambiguity. To address this challenge, we propose a novel nested NER method based on an IoU-aware denoising diffusion model, which formulates the nested NER task as a generative denoising process that progressively recovers gold entity spans from noisy span proposals. We generate noisy samples during training by gradually adding Gaussian noise to the ground-truth entity boundaries. We then train a denoiser incorporating a top-k selective attention mechanism to refine entity span proposals iteratively. To strengthen the alignment between boundary localization and entity classification, we introduce an IoU-aware loss function that optimizes the overlap between predicted and ground-truth spans. This design more accurately guides boundary regression and effectively reduces misalignment caused by conventional regression losses. Our model leverages sentence features and timesteps as conditional inputs to capture contextual information throughout the denoising process. During inference, the model generates final entity predictions by starting from random noise spans and iteratively refining them through a multi-step reverse diffusion process. We conduct extensive experiments on four nested NER datasets, ACE2004, ACE2005, GENIA, and KBP2017, as well as two flat NER datasets, CoNLL2003 and OntoNotes. Experimental results show that the proposed method consistently outperforms existing advanced models across all benchmarks, demonstrating its effectiveness.

Legal Judgment Prediction (LJP) is a research hotspot in legal intelligence, which aim to predict the judgment result based on the fact description. The existing research mainly focuses on multiclass classification methods and single-label learning by analyzing case facts, but neglect the semantic correlation between the fact descriptions and legal keyword labels, and the significance of legal keywords, which leads to unsatisfactory judicial prediction results. To address this limitation, we introduce a novel framework for LJP that enhances the utilization of legal concept keywords through prompt learning. We propose a method based on legal charge keywords and prompt engineering to enhance the performance for LJP. Our approach first integrates legal keywords with fact descriptions to improve the representation capacity of case fact vectors. We have incorporated the legal keywords into the language model to enhance the model’s ability to understand and process legal texts. And we design a prompt template to guide the reasoning process of the pre-trained language model through structured instructions. This strengthens the semantical relevance between the fact description and legal labels, so that the model can more accurately capture the logical connection between fact descriptions and charges. Experimental results on CAIL2018 datasets across different tasks show an improvement in F1 scores ranging from at least 1.59

To address the challenge of detecting small targets caused by the small size and high quantity of targets in current unmanned aerial vehicle (UAV) aerial images, we propose a novel multi-scale self-attention graph neural network model based on YOLOv8. This model can learn the weak semantic information generated between small objects, guiding the network to estimate reliable details of small objects and capture relationships between them, thereby achieving accurate detection and classification of small objects through enhancing similar features. Additionally, we introduce an improved objective box loss function to tackle the issue of high-density object detection. We evaluate our proposed model on the widely-used open-source dataset Visdrone2019 and DOTAv2. Experimental results demonstrate that our model outperforms the existing baseline YOLOv8, achieving a significant improvement of 7.40% in terms of mAP 50 . Ablation experiments further validate the effectiveness of our designed modules and loss function. Furthermore, our approach can efficiently detect small objects in complex road traffic environments, contributing to the advancement of smart city development.

FederatedLearning (FL) is an innovative framework that enables workers to collaboratively train a global shared model in a decentralized manner. Instead of transferring raw data to a centralized location, workers train the shared model locally. However, participating in federated learning tasks consumes communication resources and computing power and poses privacy risks. Naturally, workers are reluctant to engage in training without reasonable rewards. Moreover, there is a risk of malicious workers submitting harmful local models to undermine the global model and gain undeserved rewards. To tackle these challenges, we propose RIFL, which can fairly motivate honest workers to participate in FL tasks and prevent malicious workers from corrupting the global shared model. We employ centered kernel alignment (CKA) to assess the similarity between the local models submitted by workers and the global model. Subsequently, we utilize a similarity clustering-based approach to identify and eliminate local models submitted by potentially malicious workers. Additionally, a reward allocation mechanism based on reputation and data contribution is designed to motivate workers with high-quality data to participate in FL tasks and prevent intermittent attackers from gaining undeserved rewards. Finally, extensive experiments on benchmark datasets show that RIFL achieves high fairness and robustness, improving global model accuracy and motivating workers with high-quality data to participate in FL tasks under non-IID and unreliable scenarios.

Background: Cryo-electron microscopy (Cryo-EM) plays an increasingly important role in the determination of the three-dimensional (3D) structure of macromolecules. In order to achieve 3D reconstruction results close to atomic resolution, 2D single-particle image classification is not only conducive to single-particle selection, but also a key step that affects 3D reconstruction. The main task is to cluster and align 2D single-grain images into non-heterogeneous groups to obtain sharper single-grain images by averaging calculations. The main difficulties are that the cryo-EM single-particle image has a low signal-to-noise ratio (SNR), cannot manually label the data, and the projection direction is random and the distribution is unknown. Therefore, in the low SNR scenario, how to obtain the characteristic information of the effective particles, improve the clustering accuracy, and thus improve the reconstruction accuracy, is a key problem in the 2D image analysis of single particles of cryo-EM. Results: Aiming at the above problems, we propose a learnable deep clustering method and a fast alignment weighted averaging method based on frequency domain space to effectively improve the class averaging results and improve the reconstruction accuracy. In particular, it is very prominent in the feature extraction and dimensionality reduction module. Compared with the classification method based on Bayesian and great likelihood, a large amount of single particle data is required to estimate the relative angle orientation of macromolecular single particles in the 3D structure, and we propose that the clustering method shows good results. Conclusions: SimcryoCluster can use the contrastive learning method to perform well in the unlabeled high-noise cryo-EM single particle image classification task, making it an important tool for cryo-EM protein structure determination.

Blood pressure prediction is a crucial tool in preventing cardiovascular-related diseases. Therefore, improving the accuracy of blood pressure prediction plays a critical role in disease prevention. Previous models for blood pressure prediction have faced challenges related to inadequate feature extraction and insignificant effective information mining. To address these issues, this paper proposes an improved Unet-based continuous blood pressure prediction method that effectively processes the spatial information of multi-scale feature maps and establishes long-term dependencies between multi-scale channels. The proposed model achieves high accuracy meeting the requirements of the AAMI standard and the BHS A Grade. Moreover, the mean absolute errors of systolic blood pressure (SBP) and diastolic blood pressure (DBP) are 3.41 mmHg and 2.58 mmHg, respectively, with standard deviations (STD) of 6.25 mmHg and 5.08 mmHg.