Hong Huang, Decheng Wu, Qiangqiang Hu, Guanghua Yu, Jinhai Yang, Jianchen Zhu, Xue Liu, Dapeng Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Pavement crack detection is a fundamental task for maintaining the stability and sustainability of transportation infrastructure systems. Previous convolutional neural network (CNN) and Transformer-based methods have achieved high accuracy in general crack detection. However, processing slender crack images collected by uncrewed aerial vehicles (UAVs) and inspection vehicles remains extremely challenging. To this end, this paper proposes a pavement crack evaluation framework using a hybrid CNN-Mamba network and an air-ground platform. Firstly, the framework employs an air-ground platform composed of UAVs and vehicles to collect pavement crack data and perform preprocessing. Secondly, a hybrid CNN-Mamba network with wavelet transform, named WTCMamba, is proposed to achieve crack segmentation. Specifically, the network consists of a lightweight encoder composed of multiple dual convolution (DC) modules, a Mamba decoder, and a contextual spatial feature propagation (CSFP) module. The key innovation of the network is that the Mamba decoder comprises multiple wavelet-guided Mamba (WGM) modules, which introduce wavelet transform to convert feature channels into the frequency domain and fuse features in the global space. Finally, a grid-based quantitative risk evaluation method and a correlation analysis method are employed to analyze the overall risk of crack clusters and the morphological features of individual cracks. In addition, WTCMamba is deployed and tested on edge computing devices, achieving 35.63 FPS. Extensive experiments demonstrate that WTCMamba achieves excellent crack detection performance using only 2.31 M parameters and outperforms 12 state-of-the-art methods.

High-sensitivity geoscience and remote sensing instruments employing large reflector antennas face a significant threat from radio frequency interferences (RFIs). In the presence of sufficiently strong RFI, the front-end components of the receiver are driven into nonlinear operation, potentially resulting in permanent damage to the receiver. In this article, we propose utilizing a wideband true-time delay active reconfigurable intelligent surface (WTTD-ARIS) to mitigate wideband RFI encroaching on the mainlobe of a large reflector antenna before it enters the RF front end of the receiver chain. The formulated optimization problem minimizes the sum of RFI power and noise power by jointly optimizing the refractive beamforming at the WTTD-ARIS and the wideband performance of the wireless channel. To tackle this high-dimensional mathematically intractable optimization problem, a two-stage scalable partition deep reinforcement learning (DRL) algorithm is proposed to reduce the computational complexity while achieving robust optimization of the coefficients of the WTTD-ARIS. First, the large-scale WTTD-ARIS is partitioned into several subarrays and a noise power minimization guided codebook design method is proposed to obtain the preliminary coefficients of each subarray. Then, a robust feature-domain partition graph attention reinforcement learning (FPGA-RL) algorithm is developed to correlate the inherent pattern of the coefficients and the dynamic environment to obtain the final coefficients. Finally, the experimental results demonstrate the effectiveness and robustness of the proposed technique to mitigate the mainlobe wideband RFI for a large reflector antenna before it enters the RF front end.

Large language models (LLMs) have proven to be very superior to conventional methods in various tasks.However, their expensive computations and high memory requirements are prohibitive for deployment.Model quantization is an effective method for reducing this overhead.The problem is that in most previous works, the quantized model was calibrated using a few samples from the training data, which might affect the generalization of the quantized LLMs to unknown cases and tasks.Hence in this work, we explore an important question: Can we design a data-free quantization method for LLMs to guarantee its generalization performance?In this work, we propose EasyQuant, a trainingfree and data-free weight-only quantization algorithm for LLMs.Our observation indicates that two factors: outliers in the weight and quantization ranges, are essential for reducing the quantization error.Therefore, in EasyQuant, we leave the outliers (less than 1%) unchanged and optimize the quantization range to reduce the reconstruction error.With these methods, we surprisingly find that EasyQuant achieves comparable performance to the original model.Since EasyQuant does not depend on any training data, the generalization performance of quantized LLMs are safely guaranteed.Moreover, EasyQuant can be implemented in parallel so that the quantized model could be attained in a few minutes even for LLMs over 100B.To our best knowledge, we are the first work that achieves comparable performance with datadependent algorithms under a data-free setting and our algorithm runs over 10 times faster than the data-dependent methods.

Large, single-dish radio telescopes with high sensitivities are ideal for detecting faint radio transients (RTs). However, single-dish radio telescopes possess a limited angular resolution, which limits their accuracy in localizing objects. In this article, we propose to improve the localization accuracy of the RT by exploring the 3-D focal field distributions (3DFFDs) of the dish reflector with a gradient-guided attentional network (GGAN). The LSTM-based attention block of the GGAN achieves the task-oriented adaptive recalibration of 3DFFD features by exploring the significant properties and spatial dependencies of 3DFFD. In addition, a gradient-guided approach is being developed to improve the attention block performance under varying incident angles. The proposed attention mechanism is applied to the convolutional neural network in order to reconstruct 3DFFDs and perceive RT positions based on the reconstructed results. Simulation results indicate that the technique can enable the precise localization of RTs. Moreover, the proposed solution improves the telescope’s instantaneous field of view (FOV) compared to a sky survey with the traditional cluster feed telescope.

The full-waveform hyperspectral light detection and ranging (FWHSL) data have been widely used in surface topography, vegetation detection, and 3-D urban terrain modeling, capable of revealing the spatial distribution of a target and more detailed spectral information in the vertical direction. However, the echo signals of a target would significantly vary between different spectral channels due to the reflectance characteristics and the uneven energy distribution of supercontinuum laser source. Especially, band channels with weak reflectance over a long distance would affect the extraction accuracy of waveform parameters, which are essential for retrieving the spatial and spectral information of targets. This article proposes a multichannel interconnection decomposition method to improve the extraction accuracy of distance and spectral information at each pulse using hyperspectral waveform data. Two experiments were conducted to verify the performance of long-distance detection of targets using FWHSL. The first experiment detected a standard whiteboard, a green leaf, and a yellow leaf at roughly 518 m. Results demonstrated a considerable improvement in ranging precision and spectral detection using the proposed method compared with using the optimal channel with the best data quality. The second experiment simultaneously detected two adjacent targets at a distance of approximately 518 m. Results presented clear superiority of adding waveform channels in terms of discovering overlapping components and retrieving accurate waveform parameters. The success rate of extracting two targets 60 cm apart was greatly increased from 47% to 73% through the multichannel interconnection waveform decomposition (MIWD) method.