Nuo Chen, Andre Lin HuiKai, Jiaying Wu, Junyi Hou, Zining Zhang, Qian Wang, Xidong Wang, Bingsheng He. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The rapid advancement of Multimodal Large Language Models (MLLMs) has led to remarkable performances across various domains. However, this progress is accompanied by a substantial surge in the resource consumption of these models. We address this pressing issue by introducing a new approach, Token Reduction using CLIP Metric (TRIM), aimed at improving the efficiency of MLLMs without sacrificing their performance. Inspired by human attention patterns in Visual Question Answering (VQA) tasks, TRIM presents a fresh perspective on the selection and reduction of image tokens. The TRIM method has been extensively tested across 12 datasets, and the results demonstrate a significant reduction in computational overhead while maintaining a consistent level of performance. This research marks a critical stride in efficient MLLM development, promoting greater accessibility and sustainability of high-performing models.

Adapting medical Large Language Models to local languages can reduce barriers to accessing healthcare services, but data scarcity remains a significant challenge, particularly for low-resource languages. To address this, we first construct a high-quality medical dataset and conduct analysis to ensure its quality. In order to leverage the generalization capability of multilingual LLMs to efficiently scale to more resource-constrained languages, we explore the internal information flow of LLMs from a multilingual perspective using Mixture of Experts (MoE) modularity. Technically, we propose a novel MoE routing method that employs language-specific experts and cross-lingual routing. Inspired by circuit theory, our routing analysis revealed a \textit{``Spread Out in the End``} information flow mechanism: while earlier layers concentrate cross-lingual information flow, the later layers exhibit language-specific divergence. This insight directly led to the development of the Post-MoE architecture, which applies sparse routing only in the later layers while maintaining dense others. Experimental results demonstrate that this approach enhances the generalization of multilingual models to other languages while preserving interpretability. Finally, to efficiently scale the model to 50 languages, we introduce the concept of \textit{language family} experts, drawing on linguistic priors, which enables scaling the number of languages without adding additional parameters.

Wentao Ge, Shunian Chen, Hardy Chen, Nuo Chen, Junying Chen, Zhihong Chen, Wenya Xie, Shuo Yan, Chenghao Zhu, Ziyue Lin, Dingjie Song, Xidong Wang, Anningzhe Gao, Zhang Zhiyi, Jianquan Li, Xiang Wan, Benyou Wang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Zhenyang Cai, Ke Ji, Xiang Wan, Benyou Wang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Xidong Wang, Guiming Chen, Song Dingjie, Zhang Zhiyi, Zhihong Chen, Qingying Xiao, Junying Chen, Feng Jiang, Jianquan Li, Xiang Wan, Benyou Wang, Haizhou Li. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

In the upcoming sixth-generation (6G) communication system, to guarantee the realtime performance and high reliability of intelligent service, artificial intelligence (AI) is indispensable. Accurate prediction of cell traffic is essential in networking, such as energy-saving, load-balancing, handover, and ensuring the quality of service (QoS) for 6G everyone-centric customized services. In the scenario of cell traffic, inter-cell correlations, such as distances and handover operations between cells can directly affect the intra-cell traffic load concurrently. In this work, we proposed a multi-hierarchical spatial-temporal graph convolution network (MH-STGCN) to predict cell traffic indicators. The hierarchical graphs explicitly import different inter-cell correlations to improve the performance by associating these correlations with intra-cell indicators and multiple graphs allow to consider different correlations simultaneously. Therefore, MH-STGCN is capable to import correlations from multiple sources simultaneously. Experiments show that MH-STGCN is more reliable and improves MAE by 8.77% and RMSE by 26.14% on average from MTGNN. MTGNN outperforms the other baseline methods conducted on our dataset.

Beamforming plays an important role in 5G Massive Multiple-Input Multiple-Output (MMIMO) communications. Optimizing beamforming configurations for 5G base stations (BSs) can substantially improve the quality of service for mobile users, and thus has great practical value. However, identifying the optimal beamforming configurations has proven to be a complex task, and is even more challenging when accounting for the unavoidable coupled influence among multiple densely deployed 5G BSs. In this paper, we propose a highly efficient deep multi-agent Bayesian optimization framework for the coordinated beamforming optimization problem that involves multiple BSs. Its core algorithm is built upon a deep ensemble of neural networks and a sample-efficient upper confidence bound (UCB) based exploration strategy. Our numerical results show that the proposed approach is highly effective in searching for optimally coordinated beamforming vectors under extremely large search space, and beats strong multi-agent reinforcement learning baselines in terms of optimization quality and sample efficiency.

To address the three main challenges in Digital Twin Network (DTN), we propose the Elastic Digital Twin Network Modeling (EDiTNetMdl) fitting in network life cycle to restrain resource occupation. The challenges our solution aims to tackle are: fulfilling differentiated requirements across Network Life Cycle (NLC) stages with a single DTN; loosening the decision delay caused by current DTNs, hampering utilization; and the prohibitive costs of fully replicating physical telecom infrastructure and systems. To tackle these challenges, EDiTNetMdl constructs DTN modeling in a hierarchical, multilayered abstraction. In this framework, Telecom Network Elements (NE) are represented as Classes and inheritance hierarchies. NE functions and connections are transformed into class methods along these inheritance chains. NLC stages are detected through probes which are specialized methods used to deduce the stage and needs. Complete data collection occurs at the bottom NE instances, while higher stages obtain well-abstracted metadata and functional descriptions. We enumerate 7 NLC use cases as inputs for numerical evaluations. Both the number of DTN Instances built and methods invoked are significantly decreased in EDiTNetMdl compared to contrasting solutions. Consequently, the resource occupation is significantly diminished, further highlighting the efficiency gains achieved with EDiTNetMdl.

With the growing demand for mobile communication services and continuous advances in communication technologies, ubiquitous coverage and Internet of Everything have become basic capabilities required for 5G, 6G and future networks. On-orbit autonomy enables satellites to perform tasks, process data, adjust parameters, update software, diagnose faults and recover functions autonomously in orbit. This enhances intelligence, efficiency and reduces operation costs and risks, especially for large low earth orbit (LEO) constellations. On-orbit autonomy is an important trend with great significance and value. Although some research and tests have been conducted as well as application requests and standardization gathered, modeling constellation autonomy from an on-orbit communication and computing perspective has received little attention. This paper reviews related standards, research, and tests as a foundation for modeling. Key technologies, networking and management are studied as major concerns in modeling. The Constellation Autonomy Modeling (CAM) framework for agile on-orbit communication and computing is proposed as a reference model and method for on-orbit autonomy.

With the development of IT infrastructures, applications and systems generate a tsunami of data that keeps growing. Traditional IT management solutions can’t keep up with volume and complexity. Artificial intelligence for IT operations (AIOps) is an extremely effective method that could simplify IT operations management and accelerate & automate problem resolution in complex modern IT environments. Alarm root cause location is an important scenario and key function of AIOps. At present, the relevant research work mainly focuses on the association mining of historical alarm data, forming alarm rules, and processing offline alarm compression. However, the practical applications require faster and more accurate root cause location of alarms, which could put forward higher requirements for its real-time performance. The online real-time alarm root cause location algorithm proposed in this paper can fully explore the relationship between alarms in the dimension of time and space, and achieve online alarm data compression through technologies such as alarm association time, alarm event division, and alarm event topology generation. Real-time accurate division of alarm events and real-time location of key alarms greatly improve the velocity and accuracy of root cause location. The algorithm has been launched on a mobile network operator's 5G network management system. With the application of the proposed algorithm, over 10,000 alarms are processed per minute, and the accuracy of the root cause of the alarm has reached 85%, which has achieved good online effects.

Mastering the knowledge about security-sensitive functions that can potentially result in bugs is valuable to detect them. However, identifying this kind of functions is not a trivial task. Introducing machine learning-based techniques to do the task is a natural choice. Unfortunately, the approach also requires considerable prior knowledge, e.g., sufficient labelled training samples. In practice, the requirement is often hard to meet.

In this paper, we consider the problem of interference mitigation in the downlink of multicell networks via base station coordination. In this paper, a simple and efficient scheme for interference management based on potential game is proposed. The main emphasis of this paper is placed on the problem of users' quality of service (QoS) in order to maximize the efficient throughput of system. Meanwhile, a pricing factor is introduced which is proportion to the co-channel interference to other base stations. Furthermore, an improved gradient projection rule with variable step size and Jacobi iterative algorithm are utilized to solve the optimization problem. Pareto optimal is verified by using "price of anarchy" as an optimize performance indicators in potential game. Simulation results show that our proposed scheme can significantly improve the performance of multicell networks.

With the sharp increasing demand of the indoor coverage and network capacity, dense femtocells are the ultimate target of the femto cellular network deployment in urban environment However, dense femtocell deployment could introduce excessive inter-femtocell interference (IFI). To address this problem, this paper proposes a distributed power self-optimization scheme for the downlink of dense femtocell networks. First, a non-cooperative power self-optimization game framework is established based on the exact potential game, which is demonstrated to converge to a pure and unique Nash Equilibrium. Then, a novel convex pricing mechanism is presented to price the transmit power of femtocells. Finally, combined with Firefly Intelligence Optimization (FIO), an effective power self-optimization algorithm with guaranteed convergence is proposed to achieve the Nash Equilibrium. With practical LTE parameters and 3GPP dual-strip femtocell model, simulation results show that the proposed scheme dramatically reduce femtocells transmit power while improving the throughput of dense femtocell networks compared to the scheme with non-linear pricing.

Due to the individualistic nature of the femtocells and the uncertainty on the number and location of these devices, self-organization techniques play a very important role in successfully deploying and managing a large femtocell tier. This paper proposes a distributed power self-optimization scheme to suppress the interference and optimize the energy efficiency. First, the non-cooperative power self- optimization game is established, which is demonstrated to converge to a pure and unique Nash Equilibrium. Then, the impact of different power constraint conditions on the equilibrium is analyzed. Finally, a distributed power self- optimization algorithm is presented to achieve the equilibrium. With practical LTE parameters and 3GPP dual-strip femtocell model, simulation results show that the proposed scheme has fast and stable convergence and improves the energy efficiency significantly compared with Iterative Water Filling (IWF) algorithm. Furthermore, two power constraint conditions manifest different performance in terms of the energy efficiency.

Interference management is one of the most important issues in two-tier femtocell networks in which a central macrocell underlaid with several femtocell hotspots. To address the cross-tier and co-tier interference, this paper introduces an energy-efficient power control algorithm via convex pricing scheme, where both circuit and transmit power have been considered. To achieve the target SINR, the macrocell protects itself by actively adapting its interference cap on the total interference they are willing to tolerate and pricing the interference from the femtocell user equipments (FUEs), while the FUE is to maximize its own throughput and keep the fairness with the other FUEs. Then, the existence and uniqueness of the equilibrium for the proposed scheme are studied. Simulation results show that the proposed scheme can improve energy efficiency of femtocell networks and mitigate the crossand co-tier interference significantly while guaranteeing the minimum QoS requirement of macrocell users.