Jiajun Zhang, Yuying Li, Zhixun Li, Xingyu Guo, Jingzhuo Wu, Leqi Zheng, Yiran Yang, Jianke Zhang, Qingbin Li, Shannan Yan, Changguo Jia, Junfei Wu, Zilei Wang, Qiang Liu, Liang Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Industrial gas leakage detection is critically important for safety and environmental protection. While infrared imaging enables detection of invisible gases, two challenges remain: existing datasets lack realistic industrial scenarios, and current methods struggle to distinguish gas plumes from background interferences or segment discontinuous gas distributions. This paper introduces a benchmark comprising an Industrial RGB-Thermal Dataset (IRTD) with gas emission and leakage data from laboratory and industrial sites. A VLM-assisted RGBThermal detection framework with a Cross-Attention based Feature Difference (CAFD) module is designed to enhance gasspecific feature differentiation by computing inter-modal feature discrepancies. Evaluations on public datasets and IRTD demonstrate state-of-the-art results.

The rapid development of the Internet has revo-lutionized our lives, providing us with an array of convenient services. However, this revolution has also led to the problem of data overload. Personalized recommendation services have been widely recognized as an effective solution to this problem by both the industrial and academic communities. Accurate prediction of click-through rate (CTR) is a critical task in personalized recommendation systems since it can improve the user’s shopping experience, ultimately increasing revenue for the platform. To achieve accurate CTR prediction, it is crucial to capture user interests. While several methods for interest modeling exist, challenges related to user interest evolution and user behavior noise need further attention. This paper proposes a novel CTR prediction model named the Evolutionary Interest Representation Network (EIRN), which learns the evolving interest representation of users from their behaviors, profiles, and environmental attributes. In the model, we introduce an advertisement correlation graph decomposition layer and a noise filter layer to enhance and purify the user behavior representation. An interest extraction layer is designed to capture sequential correlations and common characteristics between user behaviors, representing their interests. Additionally, as user interests evolve over time and their importance changes, an attention-based interest evolution layer is designed to track the evolution of user interests. Our proposed method is validated using four public datasets and an industrial dataset, showing its superiority over existing CTR prediction methods.

Inverter-based resources (IBRs) such as photovoltaic systems and wind farms are being integrated into the power grids as part of climate change efforts. Most renewable resources utilize phase-locked loop (PLL) to establish network synchronization, and operate under the grid-following (GFL) mode. However, it has been recently reported that with the increasing amount of IBRs to replace fossil-fuel based synchronous generators, the continued use of GFL will reduce the overall stability of the grid. This issue can be resolved by changing the control mode of some inverters to grid-forming (GFM), which the PLL loop is replaced by a droop control loop. Nevertheless, the question regarding the ratio of operating IBRs under GFL and GFM in maintaining system stability has yet to be answered. This paper aims to fill this gap. We first built a high-fidelity simulation model consisting of only GFM and GFL inverters followed by extracting the stability features from the time-domain current and frequency waveforms of individual inverters. Next, an entropy weight method (EWM) is proposed to build an objective evaluation model for assessing the grid stability. Numerical studies were conducted to assess the stability due to different ratios of GFL and GFM along with different grid topologies and IBR locations. Overall, this paper serves as a guide for future placement strategies of distributed renewables within power grids.

Tourism trend prediction facilitates estimation of tourism investment and revenue. Studies on tourism prediction have primarily relied on linear models and historical visitors; however, relationships between tourism trends and their factors may be nonlinear. This study constructed factors from internet search data and predicted tourism trends using a spatiotemporal framework based on the extreme gradient boosting (XGBoost) method. The study first sorted Baidu index data that is computed by weighting the search frequency. The spatial cluster analysis was conducted to incorporate spatial characteristics, and principal component analysis was further performed to identify factors. The next step derived variables using the weighted moving average method to reduce the lag effect between tourism internet search and actual behavior. We applied the proposed spatiotemporal XGBoost composite model to predict Beijing’s tourism trends. The R2 scores of the simple XGBoost model, the autoregressive integrated moving average model, the spatial XGBoost model, and the spatiotemporal XGBoost composite model were 0.517, 0.625, 0.791, and 0.940, respectively. Compared to predictions from different models, the spatiotemporal XGBoost composite model has the best prediction ability. The findings also suggest that machine learning methods may not perform well without considering spatial properties, such as spatial autocorrelation and spatial heterogeneity.

With the rapid development of cellular network systems, the operators need more experience to deal with complicated network management system and wide range of Key Performance Indicators (KPIs). There are many indicators related to each other due to the definition or communication process. But several implicit associations still exist among these KPIs. This paper proposes an approach to figure out the implicit linear relationship among indicators clearly in which a new clustering technique is used for distinguishing different relationships. Data analysis using real network data shows that the approach can well divide data into clusters, and each cluster can effectively reflect the relationship between indicators.