Multimodal Large Language Models (MLLMs) increasingly support dynamic image resolutions. However, current evaluation paradigms primarily assess semantic performance, overlooking the critical question of resolution robustness - whether performance remains stable across varying input resolutions. To address this gap, we introduce Res-Bench, a comprehensive benchmark comprising 14,400 samples across 12 resolution levels and six core capability dimensions. We designed a novel evaluation framework that goes beyond traditional accuracy metrics to capture performance stability. This framework introduces multiple robustness metrics: Spearman's correlation for assessing resolution-performance trends, and Absolute/Relative Continuous Error (ACE/RCE) for measuring performance volatility. Using these metrics, we conducted a large-scale evaluation of leading MLLMs. Our analysis encompasses: (1) model-centric and task-centric robustness examination, (2) investigation of preprocessing strategies including padding and super-resolution, and (3) exploration of fine-tuning for stability enhancement.

There is an emerging trend of using deep learning (DL) to handle complex tasks in wireless communication systems. However, recent research suggests that DL-enabled communication systems are vulnerable to adversarial attacks. Fortunately, most of these attacks are simulation-based, incapable of handling realistic channels with, e.g., multipath fading, temporal dynamics, and hardware nonlinearity, and hence lack of practicality. To this end, we present RadioShock, an over-the-air adversarial attack against wireless communication systems. Through accurate estimations of channel states for dynamic adaptations, RadioShock is made for real-world communication scenarios.We further introduce a universal compact perturbation generation algorithm, along with a new perturbation constraint strategy, aiming to achieve covert over-the-air adversarial attacks. We implement a RadioShock prototype and conduct extensive experiments using automatic modulation classification systems as a representative application scenario. The results reveal that RadioShock diminishes the accuracy of diverse models utilized in wireless communication systems by up to 52.41%, far more effective than existing simulation-based adversarial attacks in facing real-world applications.

In recent years, Transformers have achieved significant success in image fusion. These methods utilize self-attention mechanism across different spatial or channel dimensions and have demonstrated impressive performance. However, existing methods only optimize along a single dimension and struggle to simultaneously capture the complex dependencies between spatial and channel dimensions. To address this problem, we propose a novel multi-dimensional adaptive interaction transformer network, named as MAITFuse, to enhance the multilevel information expression and detail retention capabilities of images. We design a Multi-Dimensional Feature Extraction (MDFE) module to extract features across spatial and channel dimensions in parallel, and introduce a novel weighted cross-attention fusion method to integrate multi-dimensional information effectively. Experimental results show that, compared to existing fusion methods, our proposed method achieves superior fusion performance across various datasets.

Object detection plays a crucial role in various real-world applications, particularly on resource-constrained edge devices. However, existing lightweight detection models often struggle to balance accuracy and computational efficiency due to limitations in fine-grained feature extraction and effective multiscale representation. To address these challenges, we propose an efficient object detection framework that integrates a Region-Context Enhancement Module (REM) and a Self-Weighted Feature Fusion Module (SFM). The REM captures local object details while preserving global contextual information, enabling robust feature learning. The SFM further enhances detection performance by adaptively balancing multi-scale features within the feature pyramid. Experimental results demonstrate that the proposed model not only surpasses the current state-of-the-art real-time object detection models in terms of performance but also significantly reduces resource requirements, making it highly suitable for real-time edge computing applications. For instance, when evaluated on the MS COCO dataset, our model achieves a 1.9% higher AP compared to YOLOv7-Tiny (CVPR 2023), while simultaneously reducing the number of parameters by 1.93M and FLOPs by 5.6G.

Real-time object detection is a fundamental task in the field of computer vision. In object detection network design, redundant strategies are commonly employed to capture similar features across multiple channels. However, the holistic learning of feature maps in traditional redundant strategies can further increase redundancy in features. Therefore, we propose a novel Convolution-based Feature Extraction Module (CFEM) that employs a hierarchical feature extraction and fusion strategy to capture more diverse features with lower computational costs. In addition, we design the Multiple Hybrid Detection Head (MHDH), which improves computational efficiency and reduces the number of parameters of the model by applying convolutional layers with different kernel sizes on the input features. Additionally, we replace some convolutional layers in the baseline model’s Darknet-53 backbone and Global Feature Pyramid Network (GFPN) with CFEM and create an entirely new real-time object detection network by integrating the MHDH module. This network efficiently enhances and learns multi-scale features, achieving outstanding detection accuracy with relatively fewer parameters and computational costs. Experimental results indicate that the proposed model outperforms SOTA methods. Taking the MS COCO dataset as an example, compared with YOLOv7-Tiny (CVPR 2023), our model reduces the number of parameters by 0.23M and FLOPs by 4.5G, while AP increases by 2.7%.

As an important research topic in the field of intelligent network management, network traffic prediction has received extensive attention in recent years. However, the existing research mainly focuses on short-term network traffic prediction and does not consider the influence of external factors and the effective embedding of heterogeneous data sources. Effective longterm traffic prediction has become a challenging problem. To address these challenges, this paper proposes a knowledge-driven deep learning method KGASTN for spatio-temporal graphical convolutional networks for long-term traffic flow prediction with multiple factors. In the method, our innovative idea is to design a knowledge-data dual-driven feature extraction scheme that fuses the knowledge representation of external factors into a spatio-temporal graph convolutional network. The method utilizes knowledge inference and sequence similarity algorithms to realize the extraction of explicit knowledge in log data and the construction of similarity graphs between traffic data sequences; and fuses explicit knowledge and similarity relationships based on the knowledge-data dual-drive model, and ultimately constructs spatio-temporal graph convolutional networks based on the attention mechanism. We evaluate KGASTN with a heterogeneous dataset containing log data, and use several sequence prediction datasets from other application domains for additional comparison. Experimental results show that our method outperforms several state-of-the-art baselines.

Source code identifiers often contain abbreviations. Such abbreviations may reduce the readability of the source code, which in turn hinders the maintenance of the software applications. To this end, accurate and automated approaches to expanding abbreviations in source code are desirable and abbreviation expansion has been intensively investigated. However, to the best of our knowledge, most existing approaches are heuristics, and none of them has even employed deep learning techniques, let alone the most advanced large language models (LLMs). LLMs have demonstrated cutting-edge performance in various software engineering tasks, and thus it has the potential to expand abbreviations automatically. To this end, in this paper, we present the first empirical study on prompt-based usage of LLMs for abbreviation expansion. Our evaluation results on a public benchmark suggest that GPT is substantially less accurate than the state-of-the-art approach, reducing precision and recall by 28.2% and 27.8%, respectively. We manually analyzed the failed cases, and discovered the root causes for the failures: 1) Lack of contexts and 2) Inability to recognize abbreviations. In response to the first cause, we investigated the effect of various contexts and found surrounding source code is the best selection. In response to the second cause, we designed an iterative approach that identifies and explicitly marks missed abbreviations in prompts. Finally, we propose a post-condition checking to exclude incorrect expansions that violate common sense. All such measures together make LLM-based abbreviation expansion comparable to the state of the art while avoiding expensive source code parsing and deep analysis that are indispensable for state-of-the-art approaches. Our evaluation results on open-source LLMs, i.e., DeepSeek-Coder and Llama, confirm that the post-condition checking works well with various LLMs.

The integration of Time-of-Flight (TOF) information in the reconstruction process of Positron Emission Tomography (PET) improves image qualities. However, implementing the cutting-edge model-based deep learning methods for TOF-PET reconstruction is challenging due to the substantial memory requirements. In this study, we presented a novel model-based deep learning approach, LMPDNet, for TOF-PET reconstruction from list-mode data. We addressed the issue of real-time parallel computation of the projection matrix for list-mode data, and proposed an iterative model-based module that utilized a dedicated network model for list-mode data. Our experimental results indicated that the proposed LMPDNet outperformed traditional iteration-based TOF-PET list-mode reconstruction algorithms. Additionally, we compared the spatial and temporal consumption of list-mode data and sinogram data in model-based deep learning methods, demonstrating the superiority of list-mode data in model-based TOF-PET reconstruction.