Shuang Cheng, Yuhua Jiang, Zineng Zhou, Dawei Liu, Tao Wang, Linfeng Zhang, Biqing Qi, Bowen Zhou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yuhua Jiang, Shuang Cheng, Yihao Liu, Ermo Hua, Che Jiang, Weigao Sun, Yu Cheng, Feifei Gao, Biqing Qi, Bowen Zhou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Incremental object detection (IOD) aims to sequentially learn new classes, while maintaining the capability to locate and identify old ones. Prior methodologies mainly tackle catastrophic forgetting through knowledge distillation and exemplar replay, ignoring the conflict between limited model capacity and increasing knowledge. In this paper, we propose the Dynamic object Query-based DEtection TRansformer (DyQ-DETR), which incrementally expands the model representation ability to achieve stability-plasticity tradeoff. First, a new set of learnable object queries are fed into the decoder to represent new classes. Second, we propose the isolated bipartite matching for object queries in different phases, based on disentangled self-attention. Thanks to the separate supervision and computation over object queries, we further present the risk-balanced partial calibration for effective exemplar replay. Extensive experiments demonstrate that DyQ-DETR significantly surpasses the state-of-the-art methods, with limited parameter overhead. The code is available at https://github.com/THUzhangjic/DyQ-DETR.

The scaling of large language models (LLMs) has revolutionized their capabilities in various tasks, yet this growth must be matched with efficient computational strategies. The Mixture-of-Experts (MoE) architecture stands out for its ability to scale model size without significantly increasing training costs. Despite their advantages, current MoE models often display parameter inefficiency. For instance, a pre-trained MoE-based LLM with 52 billion parameters might perform comparably to a standard model with 6.7 billion. Being a crucial part of MoE, current routers in different layers independently assign tokens without leveraging historical routing information, potentially leading to suboptimal token-expert combinations and the parameter inefficiency problem. To alleviate this issue, we introduce the Layerwise Recurrent Router for Mixture-of-Experts (RMoE). RMoE leverages a Gated Recurrent Unit (GRU) to establish dependencies between routing decisions across consecutive layers. Such layerwise recurrence can be efficiently parallelly computed for input tokens and introduces negotiable costs. Our extensive empirical evaluations demonstrate that RMoE-based language models consistently outperform a spectrum of baseline models. Furthermore, RMoE integrates a novel computation stage orthogonal to existing methods, allowing seamless compatibility with other MoE architectures. Our analyses attribute RMoE's gains to its effective cross-layer information sharing, which also improves expert selection and diversity.

In this paper, we propose the IFresher framework to improve the timeliness of multi-agent mobile augmented reality (MAR) systems. Existing works have made strides in accuracy-latency trade-offs, but fail to directly address realtime task responsiveness and multi-agent contention challenges. To bridge this gap, we introduce the concept of the age of analytics information (AoAI), which quantifies the combined impact of video analytics (VA) accuracy, transmission delay, and computational efficiency. By deriving a closed-form expression for AoAI, IFresher establishes a central control mechanism that jointly optimizes bandwidth allocation and video configuration to minimize AoAI while ensuring accuracy. Due to the mixed-integer nonlinear characteristics of the problem and the fact that each agent only has local observations, the problem is reformulated into a decentralized partially observable Markov decision process (Dec-POMDP). We propose a multi-agent reinforcement learning (MARL) algorithm, named convex-embedded transformer QMIX (CTQMIX), using the centralized training and decentralized execution (CTDE) framework for agent collaboration. Specifically, the convex optimization ensures optimal bandwidth distribution, and the transformer captures temporal dependencies between observations and actions across time steps to improve decisionmaking in dynamic environments. Evaluations with real-world experiments show that the CTQMIX outperforms state-of-theart (SOTA) algorithms.

Due to the intensive computing resource requirements, real-time video analytics applications typically need to transmit video to a server. However, the transmission inevitably suffers from network bandwidth limitations and fluctuations, making it challenging to guarantee video analytics performance. In this paper, we aim to maximize video analytics accuracy while maintaining low latency and frame loss rate, and propose a joint segment and frame bitrate adaptation (JSFBA) framework for real-time video analytics, which incorporates two reinforcement learning-based algorithms to adapt to bandwidth at both the segment and frame levels. Initially, considering the effect of video encoding on video analytics, we employ a bitrate control method to design a segment-level bitrate adaptation (SLBA) algorithm with a unique reward function. Based on the historical information of the video segments, SLBA selects the appropriate bitrate for each segment. Subsequently, by leveraging the ability to generate multiple bitrates in scalable video coding (SVC), we design a frame-level bitrate adaptation (FLBA) algorithm, which adapts to bandwidth in a more fine-grained manner by determining the number of layers sent for each frame. Extensive experiments on large-scale network traces reveal that JSFBA effectively balances various video analytics performance metrics and achieves maximum utility compared to state-of-the-art solutions.

Identifying user intents and their corresponding slots is the first step in the utterance interpretation pipeline of many task-oriented conversational AI systems. A multilingual system that does not adequately address unbalanced issues may provide unsatisfactory experiences for users who communicate in low-resource languages, limiting the system’s usability. Since data collection of machine learning models for this task is time-consuming, it is desirable to make use of existing data in a high-resource language to train models in low-resource languages. However, the development of such models has largely been hindered by the lack of multilingual datasets annotated according to the same guidelines with enough languages. In this paper, we present a new Cross-Lingual Task-Oriented Dialogue (CLTOD) Dataset comprised of 19k annotated utterances in 10 low-resource languages across 12 intent types. And we propose An Adversarial Contrastive Zero-Shot Learning for Cross-Lingual (2ACL) training strategy to obtain better multi-lingual semantic representation. To the end, we utilize this dataset and other publicly available datasets to conduct a comprehensive benchmarking study. This show that our model significantly outperforms state-of-the-art baselines under both zero-shot and few-shot settings. In particular, other models can improve performance by up to 200% with fine-tuning of our data.

BACKGROUND: Microorganisms inhabit various regions of the human body and significantly contribute to numerous diseases. Predicting the associations between microbes and diseases is crucial for understanding pathogenic mechanisms and informing prevention and treatment strategies. Biological experiments to determine these associations are time-consuming and costly. Therefore, integrating deep learning with biological networks can efficiently identify potential microbe-disease associations on a large scale. METHODS: We propose an adversarial regularized autoencoder graph neural network algorithm, named Stacked Adversarial Regularization for Microbe-Disease Associations Prediction (SARMDA), for predicting associations between microbes and diseases. First, we integrate topological structural similarity and functional similarity metrics of microbes and diseases to construct a heterogeneous network. Then, utilizing an autoencoder based on GraphSAGE, we learn both the topological and attribute representations of nodes within the constructed network. Finally, we introduce an adversarial regularized autoencoder graph neural network embedding model to address the inherent limitations of traditional GraphSAGE autoencoders in capturing global information. RESULTS: Under the five-fold cross-validation on microbe-disease pairs, SARMDA was compared with eight advanced methods using the Human Microbe-Disease Association Database (HMDAD) and Disbiome databases. The best area under the ROC curve (AUC) achieved by SARMDA on HMDAD was 0.9891$\pm$0.0057, and the best area under the precision-recall curve (AUPR) was 0.9902$\pm$0.0128. On the Disbiome dataset, the AUC was 0.9328$\pm$0.0072, and the best AUPR was 0.9233$\pm$0.0089, outperforming the other eight MDAs prediction methods. Furthermore, the effectiveness of our model was demonstrated through a detailed analysis of asthma and inflammatory bowel disease cases.

The adoption of grid-connected electric vehicles (GEVs) brings a bright prospect for promoting renewable energy. An efficient vehicle-to-grid (V2G) scheduling scheme that can deal with renewable energy volatility and protect vehicle batteries from fast aging is indispensable to enable this benefit. This article develops a novel V2G scheduling method for consuming local renewable energy in microgrids by using a mixed learning framework. It is the first attempt to integrate battery protective targets in GEVs charging management in renewable energy systems. Battery safeguard strategies are derived via an offline soft-run scheduling process, where V2G management is modeled as a constrained optimization problem based on estimated microgrid and GEVs states. Meanwhile, an online V2G regulator is built to facilitate the real-time scheduling of GEVs' charging. The extreme learning machine (ELM) algorithm is used to train the established online regulator by learning rules from soft-run strategies. The online charging coordination of GEVs is realized by the ELM regulator based on real-time sampled microgrid frequency. The effectiveness of the developed models is verified on a U.K. microgrid with actual energy generation and consumption data. This article can effectively enable V2G to promote local renewable energy with battery aging mitigated, thus economically benefiting EV owns and microgrid operators, and facilitating decarbonization at low costs.

Grid-connected electric vehicles (GEVs) and energy-transportation nexus bring a bright prospect to improve the penetration of renewable energy and the economy of microgrids (MGs). However, it is challenging to determine optimal vehicle-to-grid (V2G) strategies due to the complex battery aging mechanism and volatile MG states. This article develops a novel online battery anti-aging energy management method for energy-transportation nexus by using a novel deep reinforcement learning (DRL) framework. Based on battery aging characteristic analysis and rain-flow cycle counting technology, the quantification of aging cost in V2G strategies is realized by modeling the impact of number of cycles, depth of discharge, and charge and discharge rate. The established life loss model is used to evaluate battery anti-aging effectiveness of agent actions. The coordination of GEVs charging is modeled as multiobjective learning by using a DRL algorithm. The training objective is to maximize renewable penetration while reducing MG power fluctuations and vehicle battery aging costs. The developed energy-transportation nexus energy management method is verified to be effective in optimal power balancing and battery anti-aging control on a MG in the U.K. This article provides an efficient and economical tool for MG power balancing by optimally coordinating GEVs charging and renewable energy, thus helping promote a low-cost decarbonization transition.

In this paper, we investigate the task scheduling problem in the UAV-enable Mobile Edge-Computing (MEC) system with the objectives of minimizing the cost and the completion time. A NSGA-II algorithm is proposed for the problem under study. The solution is represented as a two-dimension location sequence. Major components of NSGA-II are delicately designed including the feasible solution generation method (FSGM) and genetic operations of crossover, mutation and selection. Three strategies are introduced in FSGM. A simulated annealing local search is integrated into the crossover operation, and meanwhile two novel mutation methods are proposed. The Pareto-based metrics are introduced to evaluate the performance of the compared algorithms. Experimental results show that the proposal is more effective and robust than the three existing algorithms.

Since drug-target data have neither class labels nor the cluster number information, they are not suitable for clustering algorithms that require predefined parameters determined by comparing clustering results with real class labels. Density peaks clustering (DPC) is a density-based clustering alg orithm that can determine the number of clusters without requiring class labels. However, the predefined cutoff distance of local density limits its wide application. Therefore, this paper proposes an improved local density method based on a cutoff distance sequence that overcomes the limitations of DPC and can be successful applied to drug-target data. We also introduce multiple-dimensional scaling based on drug and target similarity and perform intuitive graph analysis of the two most significant differentiation features. Drugs of the Enzyme, GPCR, Ion Channel, and Nuclear Receptor 4 standard datasets are identified as 6, 6, 3, and 5 clusters by an improved algorithm, respectively, and similarly, their targets are identified be 5, 5, 8, and 4 clusters. Drug-target data clustering results of the improved algorithm are more reasonable than the results of the fast K-medoids and hierarchical clustering algorithms.

MicroRNAs (miRNAs) regulate genes that are associated with various diseases. To better understand miRNAs, the miRNA regulatory mechanism needs to be investigated and the real targets identified. Here, we present miRTDL, a new miRNA target prediction algorithm based on convolutional neural network (CNN). The CNN automatically extracts essential information from the input data rather than completely relying on the input dataset generated artificially when the precise miRNA target mechanisms are poorly known. In this work, the constraint relaxing method is first used to construct a balanced training dataset to avoid inaccurate predictions caused by the existing unbalanced dataset. The miRTDL is then applied to 1,606 experimentally validated miRNA target pairs. Finally, the results show that our miRTDL outperforms the existing target prediction algorithms and achieves significantly higher sensitivity, specificity and accuracy of 88.43, 96.44, and 89.98 percent, respectively. We also investigate the miRNA target mechanism, and the results show that the complementation features are more important than the others.

MicroRNAs (miRNAs) are small non-coding RNAs which cause target genes degradation or translational inhibition. Constructing functional miRNAs regulatory module can be a significant step towards the discovery of their regulatory roles in various development programs. In this paper, we present a Correlated Correspondence Regulatory Module model which builds on modified Correlated Topic Model (CTM). We apply the proposed method to the expression profiles of miRNAs and genes on 89 human cancer samples. The approach computationally predicts miRNA-gene interactions according to the negative or positive correlation relationship between miRNA and gene expression data and identifies functional miRNA regulatory modules from which we can infer multiple and dynamic miRNA function according to the known elements, the result shows consistency with published literature and database. Furthermore, a miRNA regulatory network is constructed in order to study the associations among various regulatory modules, we can detect evolution of miRNA function in biological process according these associations, which solve restriction of traditional methods that only focus on static miRNA function in single regulatory module. Online services can be accessed at the website (http://nclab.hit.edu.cn/CCRM).