Xianzhen Luo, Wenzhen Zheng, Qingfu Zhu, Rongyi Zhang, Houyi Li, Siming Huang, YuanTao Fan, Wanxiang Che. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Siming Huang, Tianhao Cheng, Jason Klein Liu, Weidi Xu, Jiaran Hao, Liuyihan Song, Yang Xu, Jian Yang, Jiaheng Liu, Chenchen Zhang, Linzheng Chai, Ruifeng Yuan, Xianzhen Luo, Qiufeng Wang, YuanTao Fan, Qingfu Zhu, Zhaoxiang Zhang, Yang Gao, Jie Fu, Qian Liu, Houyi Li, Ge Zhang, Yuan Qi, Xu Yinghui, Wei Chu, Zili Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

ABSTRACT Over the years, object detection technology based on deep learning has attracted extensive research in autonomous driving. Achieving a robust object detection network under all weather conditions (e.g., sunny, fog, nighttime, rain and snow) is highly significant for autonomous driving systems, which ensure safety by recognising pedestrians, vehicles, traffic lights, etc. This paper proposes a robust multi‐object detection network named All Weather‐You Only Look Once (AW‐YOLO) based on YOLOv8, with a trade‐off between precision and lightweightness. Considering the blurring or absence of the salient object features of the image under all weather conditions, we propose a developed dilation‐wise residual (D‐DWR) module. Specifically, it combines the dilatation‐wise residual module with the dilated re‐param block using a large kernel convolution to see wide without going deep, greatly improving the feature extraction ability. Moreover, we introduce an efficient dynamic upsampler (DySample) that formulates upsampling from the viewpoint of point sampling and avoids dynamic convolution, which can improve the network's ability to feature fusion. Lightweight is an essential requirement for autonomous driving. To this end, we adopt a multi‐scale shared detection head (MSSD‐Head) to achieve lightweight deployment in autonomous vehicles. Experimental results show that the mAP50‐95 values of AW‐YOLO on the KITTI and Adverse Conditions Dataset with Correspondences (ACDC) datasets exceed the baseline model YOLOv8 by 1.7% and 1.5%, respectively. Meanwhile, the parameters and model size of AW‐YOLO have decreased by 21.4% and 20.4%, respectively.

False data injection attacks (FDIAs) refer to attackers exploiting vulnerabilities in the detection of bad data in smart grid energy management systems to maliciously manipulate state estimation results in the cyber-physical system, resulting in unstable operation of the power system. Existing deep learning-based detection schemes often fail to capture the spatial topology features and long-term dependencies in power grid data well. Meanwhile, the complexity of deep detection models makes them a "black box", reducing the credibility of detection results. To address the aforementioned challenges, this paper presents a dependency-aware deep interpretable FDIA detection model. The proposed model firstly introduces Graph Sample and Aggregate (GraphSAGE) network to extract spatial topological features, which are used to represent the deep spatial topological dependencies of adjacent data nodes. Subsequently, we build a Bidirectional Long Short-Term Memory (BiLSTM) network with a Squeeze-and-Excitation (SE) attention module, which can efficiently aggregate attack characteristics and long-term dependency information by dynamically capturing the potential correlations between FDIAs detection and measurement data. Furthermore, the SHapley Additive exPlanations (SHAP) method is used to demonstrate the interpretability of the model in the spatial-temporal dimensions and then provide the basis for high-precision detection results. A series of extensive experiments are carried out over the IEEE 14-bus and 118-bus test systems. The experimental results demonstrate that the proposed model presents a superior overall performance comparing with several state-of-the-art FDIA detection models, and provides reasonable interpretability from the spatial-temporal dimensions.

Recently, transformer-based methods have achieved state-of-the-art prediction quality on human pose estimation(HPE). Nonetheless, most of these top-performing transformer-based models are too computation-consuming and storage-demanding to deploy on edge computing platforms. Those transformer-based models that require fewer resources are prone to under-fitting due to their smaller scale and thus perform notably worse than their larger counterparts. Given this conundrum, we introduce SD-Pose, a new self-distillation method for improving the performance of small transformer-based models. To mitigate the problem of under-fitting, we design a transformer module named Multi-Cycled Transformer(MCT) based on multiple-cycled forwards to more fully exploit the potential of small model parameters. Further, in order to prevent the additional inference compute-consuming brought by MCT, we introduce a self-distillation scheme, extracting the knowledge from the MCT module to a naive forward model. Specifically, on the MSCOCO validation dataset, SDPose-T obtains 69.7% mAP with 4.4M parameters and 1.8 GFLOPs. Furthermore, SDPose-S-V2 obtains 73.5% mAP on the MSCOCO validation dataset with 6.2M parameters and 4.7 GFLOPs, achieving a new state-of-the-art among predominant tiny neural network methods.