Multimodal Large Language Models (MLLMs) have recently achieved strong performance across a variety of multimodal tasks. However, they still suffer from various forms of hallucination, which hinder their practical deployment. Prior approaches often struggle to efficiently construct high-quality hallucination-related samples and to process them in a fine-grained manner, resulting in limited effectiveness in hallucination alleviation. To address this issue, we propose a data sampling strategy that selects samples better suited for hallucination-oriented training, thereby enhancing training effectiveness. In addition, we introduce a quantitative method for measuring hallucination severity and assign individualized weights to training samples accordingly. Building on this, we present Hallucination-Differentiated Direct Preference Optimization (HD-DPO), a novel preference optimization framework. During fine-tuning, HD-DPO incorporates these weights into both the formulation of customized loss functions and the modulation of localized visual attention, enabling fine-grained optimization. Experimental results demonstrate that our method outperforms existing fine-tuning strategies across multiple benchmarks and generalizes well to diverse MLLM architectures, effectively reducing hallucination rates and enhancing overall model performance.

Whole Slide Images (WSI) play a vital role in disease diagnosis, especially for complex diseases like tumors. Although deep learning, particularly Multi-Instance Learning, has advanced pathological image analysis, challenges remain: (1) the high resolution of WSIs and limited annotations hinder supervised learning; (2) fixed data distributions lead to overfitting and poor generalization. To address these issues, we propose BTDA-MIL, a novel MIL model based on Bag Transformation Data Augmentation. BTDA-MIL designs two transformation strategies-bag merging and bag sampling-to dynamically expand and diversify training data. The model adopts a teacher-student framework, where the teacher's learned knowledge guides pseudo-bag construction, enhancing the student's classification ability. This dynamic augmentation exposes the model to varying data distributions, reducing overfitting risks. Experiments on CPTAC and TCGA-RCC datasets demonstrate that BTDA-MIL outperforms state-of-the-art MIL methods, achieving superior WSI classification performance.

Accurate interpretation and visualization of human instructions are crucial for text-to-image (T2I) synthesis. However, current models struggle to capture semantic variations from word order changes, and existing evaluations, relying on indirect metrics like text-image similarity, fail to reliably assess these challenges. This often obscures poor performance on complex or uncommon linguistic patterns by the focus on frequent word combinations. To address these deficiencies, we propose a novel metric called SemVarEffect and a benchmark named SemVarBench, designed to evaluate the causality between semantic variations in inputs and outputs in T2I synthesis. Semantic variations are achieved through two types of linguistic permutations, while avoiding easily predictable literal variations. Experiments reveal that the CogView-3-Plus and Ideogram 2 performed the best, achieving a score of 0.2/1. Semantic variations in object relations are less understood than attributes, scoring 0.07/1 compared to 0.17-0.19/1. We found that cross-modal alignment in UNet or Transformers plays a crucial role in handling semantic variations, a factor previously overlooked by a focus on textual encoders. Our work establishes an effective evaluation framework that advances the T2I synthesis community's exploration of human instruction understanding. Our benchmark and code are available at https://github.com/zhuxiangru/SemVarBench.

In the sixth generation (6G) networks, empowered by artificial intelligence/machine learning (AI/ML) techniques, the Internet of Things (IoT) is of paramount importance to realize ubiquitous connectivity for massive devices. As an open and flexible architecture, Open Radio Access Network (O-RAN) has enabled the integration of a variety of AI/ML paradigms, including federated learning. However, due to the heterogeneity of IoT devices, the generalization capability of federated learning models and energy efficiency in O-RAN architecture cannot be well guaranteed. In order to tackle these challenges, a hierarchical federated learning paradigm is proposed for resource-constrained IoT devices. First, a client-edge-cloud collaboration framework in O-RAN is designed, in which the cell-level federated learning models generated on the edge are integrated on the regional cloud based on ensemble learning method for adapting to diverse inference data streams of IoT devices. Second, the generalization performance of the framework is analyzed in terms of inference accuracy gap. A tractable upper bound can be obtained, based on which an algorithm is designed to minimize the gap by jointly optimizing on device scheduling, resource allocation and power control until approaching an optimal solution. Finally, the simulation results are provided to verify the performance gains in enhancing inference accuracy and reducing power consumption.

For millimeter wave, accurate beamforming can provide massive multiple-input multiple-output (MIMO) systems throughput gain. However, the time complexity of searching for the best transmitting and receiving beam pairs increases with the numbers of transmitting and receiving beams. In this work, we propose a deep learning (DL)-based spatial domain beam prediction method. Different from existing DL-based beam prediction network, we consider multiple kinds of configurations of beam pair subset for the input of the network, which can make one network support the prediction tasks of different antenna configurations with smaller input size. By considering the side information, we propose a multimodal structure to improve the prediction accuracy of DL-based beam prediction network. Compared to existing beam prediction methods, the proposed DL-based beam prediction network has better generalization and can improve training efficiency of the neural network with an outperformed beam prediction result.

The open radio access network (O-RAN) is evolving towards open and inter-operable RAN architecture by introducing the RAN Intelligent Controller (RIC) for the sixth generation (6G) communication systems. The integration of artificial intelligence/machine learning (AI/ML) models into O-RAN is critical to enabling network intelligence, in which federated learning has been proposed as an efficient paradigm. However, the generalization capability of federated learning models cannot be well guaranteed under heterogenous wireless circumstances. Therefore, to improve the generalization performance of models in O-RAN and to ensure the near real-time intelligence, in this paper, we have proposed a paradigm of clustered federated learning with model ensemble. First, the framework is provided, in which clustered federated learning is implemented for model training to improve convergence speed, and transfer learning is executed for near real-time model inference to adapt to diverse target tasks. Second, the generalization performance compared with optimal learning scheme is analyzed and an upper bound can be derived, based on which a joint optimization algorithm is sophisticatedly designed on transmit power control, user clustering and model ensemble. Finally, the proposed scheme is evaluated on the MNIST and CIFAR-10 data sets, which verifies the significant performance gains for deploying federated learning in O-RAN.

Robust tensor completion, which aims to recover a tensor from partial observations corrupted by Gaussian noise and sparse noise simultaneously, has a wide range of applications in visual data recovery. The existing approaches make use of convex or nonconvex relaxation based on transformed tensor nuclear norm, which may be challenged since only the global low-rankness of the underlying tensor data is utilized. In order to explore the global and local patterns simultaneously, in this paper, we propose a nonconvex model for robust tensor completion by combining the dictionary learning and nonconvex regularization. For the sake of exploring the global low-rankness, a family of nonconvex functions are employed onto the singular values of all frontal slices of the underlying tensor in the transformed domain. The dictionary learning is utilized to elucidate the local patterns of the underlying tensor data. Moreover, a family of nonconvex functions are used onto each entry of the sparse noise, which can obtain sparser solutions compared with tensor$\ell _{1}$norm. A proximal alternating linearized minimization algorithm is designed to solve the proposed model, whose convergence is established under very mild conditions. Extensive numerical experiments on multispectral images, video, and magnetic resonance imaging datasets show that the proposed model outperforms other state-of-the-art approaches.

Abstract Diagnosis prediction exploits electronic health records (EHRs) to predict the future diagnoses of patients, further supporting clinical decision making and personalized treatments. However, a patient's EHR is an irregular sequence of visits that contains a large number of medical concepts. The disease progression patterns are closely related to the visits, as well as the contextual knowledge of each visit. The existing diagnosis prediction methods ignore the complex relationships between the visits and the contextual knowledge, and thus cannot achieve satisfactory performance. Therefore, we develop a knowledge‐aware representation learning method to comprehensively model these complex relationships. Specifically, we first construct a medical knowledge graph to model the correlations between medical concepts in EHRs, and project the contextual knowledge into the pre‐learned vectors. We then devise an enhanced gated recurrent unit (GRU) neural network to extract the long‐term intra‐relationships between visits, and design a novel knowledge attention module to capture the complex inter‐relationships between the visits and the contextual knowledge. Armed with these, we provide a powerful and flexible framework to capture the long‐term discriminative disease progression patterns for diagnosis prediction. Intensive experiments are conducted on two real‐world EHR datasets. And the model achieves a competitive performance with Code‐level Accuracy@20 of 0.7465 and Visit‐level Precision@20 of 0.7547 on MIMIC‐III, and Code‐level Accuracy@20 of 0.9337 and Visit‐level Precision@20 of 0.9427 on MIMIC‐IV.

User representation is essential for providing high-quality commercial services in industry. In our business scenarios, we face the challenge of learning universal (general-purpose) user representation. The universal representation is expected to be informative, and can handle various types of real-world applications without fine-tuning (e.g., applicable for both user profiling and the recall process in advertising). It shows great advantages compared to the solution of training a specific model for each downstream application. Specifically, we attempt to improve universal user representation from two points of views. First, a contrastive self-supervised learning paradigm is presented to guide the representation model training. It provides a unified framework that allows for long-term or short-term interest representation learning in a data-driven manner. Moreover, a novel multi-interest extraction module is presented. The module introduces an interest dictionary to capture principal interests of the given user, and then generate his/her interest-oriented representations via behavior aggregation. Experimental results demonstrate the effectiveness and applicability of the learned user representations. Such an industrial solution has now been deployed in various real-world tasks.

BACKGROUND: Multi-label classification of data remains to be a challenging problem. Because of the complexity of the data, it is sometimes difficult to infer information about classes that are not mutually exclusive. For medical data, patients could have symptoms of multiple different diseases at the same time and it is important to develop tools that help to identify problems early. Intelligent health risk prediction models built with deep learning architectures offer a powerful tool for physicians to identify patterns in patient data that indicate risks associated with certain types of chronic diseases. RESULTS: Physical examination records of 110,300 anonymous patients were used to predict diabetes, hypertension, fatty liver, a combination of these three chronic diseases, and the absence of disease (8 classes in total). The dataset was split into training (90%) and testing (10%) sub-datasets. Ten-fold cross validation was used to evaluate prediction accuracy with metrics such as precision, recall, and F-score. Deep Learning (DL) architectures were compared with standard and state-of-the-art multi-label classification methods. Preliminary results suggest that Deep Neural Networks (DNN), a DL architecture, when applied to multi-label classification of chronic diseases, produced accuracy that was comparable to that of common methods such as Support Vector Machines. We have implemented DNNs to handle both problem transformation and algorithm adaption type multi-label methods and compare both to see which is preferable. CONCLUSIONS: Deep Learning architectures have the potential of inferring more information about the patterns of physical examination data than common classification methods. The advanced techniques of Deep Learning can be used to identify the significance of different features from physical examination data as well as to learn the contributions of each feature that impact a patient's risk for chronic diseases. However, accurate prediction of chronic disease risks remains a challenging problem that warrants further studies.

Massive multiple-input multiple-output (MIMO) system, in which base stations are equipped with a large number of antennas in a two-dimensional (2D) antenna array, is one of the most promising technologies to improve the spectral efficiency of cellular networks. Since there is no channel reciprocity, the frequency-division duplexing (FDD) massive MIMO system has to obtain channel state information (CSI) with the help of uplink feedback. However, in practice, there are several challenges for the implementation of FDD massive MIMO system, such as feedback overhead and delay. In this paper, we begin with the virtual sectorization using CSI- Reference Signal (RS) beamforming scheme to illustrate the implementation of FDD massive MIMO in the standard transparent manner, and then we describe the potential performance degradation caused by feedback delay. To cope with this issue, we propose a historical precoding matrix indicator (PMI) based multi-user scheduling algorithm to enhance the system performance. By utilizing system-level simulator, we further present the effectiveness of the proposed algorithms.

Scheduling with single carrier property restriction has a significant impact on system performance in Long Term Evolution (LTE) uplink (UL). Due to the unavoidable delay of control signaling, the scheduling decisions have great influence on the variation of inter-cell interference (ICI) which affects the performance of adaptive modulation and coding (AMC). Many studies have been carried out on the topic of resource allocation in single-carrier frequency division multiple access (SC-FDMA) system. Some of them deal with the ICI variation with inter-cell measurements or coordination. In this paper, we present the problem of frequency domain packet scheduling (FDPS), analyze the negative effects of ICI variation and develop our improved proportional fair (PF) scheduling algorithm based on the traditional one for LTE UL. Our proposed algorithm decreases the ICI variation and improves AMC accuracy without any inter-cell coordination. Compared with the traditional way, system level performance shows that at least 29% gain can be obtained at both the cell user average spectral efficiency and the cell edge user spectral efficiency with comparable proportional fairness in typical interference-limited scenario.

Adaptive equalization technology requires a long training sequence to update the parameters of the taps by gradient descent method step by step. In order to decrease the number of training sequence, this paper proposes an improved version of the classical differential evolution algorithm for adaptive equalizer to estimate the parameters, in which two trial vectors are created by crossover operator. The modified algorithm speeds up the convergence rate and improves the convergence precision through the evolution of multi-generation in the situation of a short training set. Compared with the traditional least mean squares (LMS) algorithm and the classical differential evolution (CDE) algorithm, the modified algorithm can switch to data transmission mode from the training mode much earlier; at the same time improve the efficiency of the transmission greatly. The simulation results have confirmed that the proposed algorithm achieves the faster convergence rate, the lower misadjustment and the less symbol error rate than the LMS algorithm and CDE algorithm in 4-PAM and 16-QAM signal systems.