Tianyu Zong, Rui Dai, Hongzhu Yi, Yuanxiang Wang, Zhenghao Zhang, Zhenyu Guan, Yujia Yang, Bingkang Shi, Yueyang Ding, Xiangxiang Chu, Kaikui Liu, Jungang Xu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multi-agent collaborative perception enhances perception performance by enabling information sharing and complementary data fusion among agents. However, this process inevitably requires a trade-off between perception performance, computing resources, and communication bandwidth to ensure overall collaboration efficiency. To address these challenges, we propose a joint resource and intelligence optimization method for multi-vehicle collaborative perception, named C3I-JO. Based on slimmable network and accuracy-awareness, the method performs joint optimization over collaborative mode, resource allocation, and intelligent elasticity. C3I-JO minimizes overall resource consumption while satisfying both perception accuracy and delay constraints, thereby improving the overall system efficiency. Simulation results demonstrate that, compared with baseline methods, the proposed method achieves superior performance in terms of both resource consumption and perception quality.

Constraint handling during tracking operations is at the core of many real-world control implementations and is well understood when dynamic models of the underlying system exist, yet becomes more challenging when data-driven models are used to describe the nonlinear system at hand. We seek to combine the nonlinear modeling capabilities of a wide class of neural networks with the constraint-handling guarantees of model predictive control in a rigorous and online computationally tractable framework. The class of networks considered can be captured using Koopman operators, and are integrated into a Koopman-based predictive tracking control (KPTC) for nonlinear systems to track piecewise constant references. The effect of model mismatch between original nonlinear dynamics and its trained Koopman linear model is handled by using a constraint-tightening approach in the proposed KPTC controller. By choosing two Lyapunov functions, we prove that the solution is recursively feasible and input-to-state stable to a neighborhood of both online and offline optimal reachable steady outputs in the presence of bounded modeling errors under certain assumptions. The proposed approach has the advantage relative to existing model-based tracking approaches of enabling data-driven models to be utilized with explicit guarantees, while using efficient quadratic program solvers in online implementations. We demonstrate the proposed approach initially in simulations, and then experimentally to the problem of reference tracking by an autonomous ground vehicle.

Collaborative perception enables connected and autonomous vehicles (CAVs) to overcome individual viewpoint limitations by exchanging perception data, making effective resource allocation crucial for timely transmission. However, existing studies focus on resource allocation within a single collaborative perception group (CPG), limiting their effectiveness in multi-group collaborative perception systems. In such a system, each CPG contributes differently to the overall collaborative perception performance, and it is challenging to evaluate and represent CPG system-level utilities. Meanwhile, competition for shared spectrum resources leads to interference and complicates the joint optimization of collaboration mechanisms and spectrum allocation, which is intensified by their temporal scale misalignment. To address these challenges, we propose a Utility-Aware Hierarchical Reinforcement Learning method (UAHRL) to jointly optimize collaboration mechanisms and spectrum allocation. Specifically, we introduce a hierarchical framework to handle temporally misaligned decisions through joint training. The upper layer optimizes the collaborative relationship and granularity over a longer time scale to enhance system-level collaborative performance, while the lower layer allocates spectrum resources over a shorter time interval to fulfill individual CPG transmission demand and enhance system transmission efficiency. To represent and utilize system-level utility, we leverage a feature-based confidence map to assess CAVs’ perception capability and complementarity. A mixing network in the upper layer further decomposes global performance into individual CPG utilities, enabling utility-aware resource allocation. Simulations show that UAHRL outperforms baseline methods in system-level collaborative perception in multi-group systems.

Because of the primary role of land surface temperature (LST) in the physical processes of surface energy balance at local through global scales, dynamic, continuous, and seamless LST monitoring is constantly in urgent need. Thermal infrared (TIR) remote sensing serves as the most commonly used sources for LST retrieval owing to its relatively fine spatial-temporal resolution and presentable accuracy. However, limited by the inability to penetrate clouds, original TIR LST data suffers significantly from data missing problems. Furthermore, the view time of pixels along the scan line differs significantly for polar-orbiting satellites, exerting appreciable influence on the subsequent data applications. To cope with the above setbacks simultaneously, we proposed a practical reconstruction framework based on the inner physical connection between LST and solar radiation, which was accurately expressed by random forest regression model, with the consideration of various auxiliary environmental factors (i.e., elevation, slope, longitude, latitude, and surface reflectance). Taking the Tibetan Plateau (TP) as the study area, the proposed method was applied to generate spatially seamless and time-consistent LST products with the use of the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra daytime LST product. From visual assessment, the reconstructed product exhibits ideal spatial-temporal continuity within the TP. Through the validation with in-situ observations from five different stations, the results show a higher consistency with ground measurements than the LST product from the Global Land Data Assimilation System (GLDAS) and other all-weather LST product, with an average improvement on RMSE of 1.06 K and 1.59 K under clear conditions, and 1.86 K and 2.72 K under cloudy conditions. The validation demonstrates that the proposed method is well applicable for all-weather LST reconstruction over a large-scale area with significant surface heterogeneity, which also shows good ability to remove the temporal inconsistency induced by satellite observations. Additionally, it can be reliably generalized to different areas with similar data requirements for its sufficient effectiveness and flexibility.

In the face of rapid global climate change and increasing occurrence of extreme weather events, acquiring seamless land surface temperature (LST) with high spatial and temporal resolution on a global scale has become increasingly crucial. However, the limited ability of Thermal Infrared (TIR) Remote Sensing to penetrate cloud cover has hindered the widespread application of TIR LST datasets. To address this limitation, we propose a novel reconstruction approach for cloud-covered pixels, which is established based on the annual surface temperature cycle. It shifted previous reconstruction from directly modelling LST to indirectly modelling the residual term derived from the LST observations and the annual surface temperature cycle (ATC) model fitted values. A random forest regression was used to build this estimation model and the model was applied to cloud-covered pixels to derive their LSTs. Taking the Iberia Peninsula as the study area, the proposed method was applied to generate the all-weather LST product of whole year 2021. The visual assessment demonstrates its robust performance across different seasons and weather conditions. Additionally, through the validation with the masked clear-sky LST observations, it reveals that the proposed method achieves a stable estimation accuracy, with the average value of the coefficient of determination (R2) and Root Mean Squared Error (RMSE) of above 0.8 and 1.08 K under different climatic conditions. In comparison, the validation with the ERA-5 land reanalysis data also indicates a relatively good consistency between the performance of the reconstructed LST and the clear-sky LST, although with a slight decline in R2and RMSE. Additionally, the indirect validation with near surface air temperature (NSAT) also shows the comparable ability of the reconstructed LST in NSAT estimation as the clear-sky LST, with an increase of RMSE no more than 0.95 K. In general, the proposed method shows good potentials in reconstructing cloud-covered LSTs with relatively stable performance under different cloud cover conditions and it can be applied for generating all-weather LST product.

Aiming to extrapolate missing facts in the future timestamp, temporal knowledge graph (TKG) reasoning has significant practical value across various applications. Most existing methods mainly rely on direct historical interactions to extract the temporal features of entities and have achieved promising performance in some specific scenes. However, these methods are constrained when it comes to predicting emerging events that lack such direct historical interaction information. To address this limitation, we propose a novel approach called the Homogeneous Entity Context Enhanced Representation Network (HECERN). Homogeneous entities are defined as entities with similar behavior patterns. HECERN takes advantage of the complex interaction connections among the relations of homogeneous entities, which contain rich and relevant context information for accurate prediction of emerging events. Specifically, we generate corresponding entity context subgraph for each history subgraph and extract dependencies for homogeneous entities and neighboring entities based on these two subgraphs, respectively. In addition, we introduce an attention-guided fusion mechanism that dynamically integrates information from neighbor-level dependencies and homogeneous-level dependencies to effectively generate the final entity representation. The effectiveness of HECERN is demonstrated through comprehensive experiments on three publicly available event-based TKG datasets. The results clearly indicate that our proposed approach significantly outperforms state-of-the-art methods in accurately predicting future events. (The source code is available at https://anonymous.4open.science/r/HECERN-2C6B.)

Land surface temperature (LST) is an important parameter in land surface processes with strong relationship between surface energy and water exchange. To effectively capture the surface thermal dynamics, the annual temperature cycle model is a good option by depicting the annual variation as a constant term plus a sine function. However, this type model suffers from the assumption of constant surface thermal property which is hardly satisfied due to the changes in vegetation cover. To well address this issue, the normalized difference vegetation index (NDVI) is introduced as an indicator to characterize the variation in surface thermal property and added to the original form to propose an improved version. Through comparison between the fitting effects of the proposed model with the original one, the improvement shows good performance in suppressing the annual maximum temperature and elevating the annual minimum temperature with the increase in vegetation cover. The difference in the annual maximum and minimum temperature between the estimates from the proposed model and the original model shows good linear regression with NDVI difference when compared with the annual mean value, with the speed of −3.22 and 4.84, respectively. In addition, the fitting accuracy is also improved with a slight increase in the coefficient of determination (0.002) and a decrease in the root mean squared error (0.018 K). The application of the proposed model also provides reasonable distribution of the annual temperature parameters in the southwest of Europe and part of North Africa, confirming its potential effect in thermal dynamic monitoring.

Recently, optical remote-sensing images have been steadily growing in size, as they contain massive data and complex backgrounds. This trend presents several problems for object detection, for example, increased computation time and memory consumption and more false positives due to the complex backgrounds of large-scale images. Inspired by deep neural networks combined with time-frequency analysis, we propose a time-frequency analysis-based object detection method for large-scale remote-sensing images with complex backgrounds. We utilize wavelet decomposition to carry out a time-frequency transform and then integrate it with deep learning in feature optimization. To effectively capture the time-frequency features, we propose a feature optimization method based on deep reinforcement learning to select the dominant time-frequency channels. Furthermore, we design a discrete wavelet multiscale attention mechanism (DW-MAM), enabling the detector to concentrate on the object area rather than the background. Extensive experiments show that the proposed method of learning from time-frequency channels not only solves the challenges of large-scale and complex backgrounds, but also improves the performance compared to the original state-of-the-art object detection methods. In addition, the proposed method can be used with almost all object detection neural networks, regardless of whether they are anchor-based or anchor-free detectors, horizontal or rotation detectors.