Chen Xu, Yu ji, Zhenyu Lv, Yang Yi, Yizhe Yang, Luyao Ji, Chaoyi Chen, Xianyang Wang, Tian Lan, Zhihua Wang, Juan Wang, Xunde Dong, Fuze Tian, Qunxi Dong, Bin Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Controlling mixed platoons, which consist of both connected and automated vehicles (CAVs) and human-driven vehicles (HDVs), poses significant challenges due to the uncertain and unknown human driving behaviors. Data-driven control methods offer promising solutions by leveraging available trajectory data, but their performance can be compromised by noise and attacks. To address this issue, this paper proposes a Robust Data-EnablEd Predictive Leading Cruise Control (RDeeP-LCC) framework based on data-driven reachability analysis. The framework over-approximates system dynamics under noise and attack using a matrix zonotope set derived from data, and develops a stabilizing feedback control law. By decoupling the mixed platoon system into nominal and error components, we employ data-driven reachability sets to recursively compute error reachable sets that account for noise and attacks, and obtain tightened safety constraints of the nominal system. This leads to a robust data-driven predictive control framework, solved in a tube-based control manner. Human-in-the-loop experiments demonstrate that theRDeeP-LCCmethod significantly improves robustness against noise and attacks, while enhancing tracking accuracy, control efficiency, energy economy, driving comfort, and driving safety.

Graph-based Retrieval-Augmented Generation (GraphRAG) has emerged as a promising paradigm for enhancing LLM reliability by enabling multi-hop reasoning over graph-structured knowledge. However, existing LLMs struggle to efficiently process graph-structured inputs, as traditional attention mechanisms are sequence-based and introduce significant redundancy when serializing graphs into prompt sequences, leading to excessive computation and memory overhead. To address this, we introduce dependency attention, a novel graph-aware attention mechanism that restricts attention computation to token pairs with structural dependencies in the retrieved subgraph. Unlike standard self-attention that computes fully connected interactions, dependency attention prunes irrelevant token pairs and reuses computations along shared relational paths, substantially reducing inference overhead. Building on this idea, we develop DepCache, a KV cache management framework tailored for dependency attention. DepCache enables efficient KV cache reuse through (i) a graph-based KV cache reuse strategy that aligns KV caches across varying prompt contexts, enabling efficient cross-request reuse in GraphRAG, and (ii) a locality-aware replacement policy that leverages spatial and temporal access patterns to improve KV cache hit rate. Evaluations across diverse models and datasets show that DepCache improves LLM inference throughput by 1.5×-5.0× and reduces time-to-first-token latency by up to 3.2×, without compromising generation accuracy.

Graph neural networks (GNNs) have emerged as a promising method for learning from graph data, but large-scale GNN training requires extensive memory and computation resources. To address this, researchers have proposed using multi-GPU processing, which partitions graph data across GPUs for parallel training. However, vertex dependencies in multi-GPU GNN training lead to significant neighbor replications across GPUs, increasing memory consumption. The substantial intermediate data generated during training further exacerbates this issue. Neighbor replication and intermediate data constitute the primary memory consumption in GNN training (i.e., typically accounting for over 80%). In this work, we propose GNN task parallelism for multi-GPU GNN training, which reduces neighbor replication by partitioning training tasks in each layer across different GPUs rather than partitioning the graph structure. This approach only partitions the graph data within individual GPUs, reducing the memory requirements of single tasks while overlapping subgraph computation across different GPUs. Shared neighbor embeddings among different subgraphs can be efficiently reused within a single GPU. Additionally, we employ a task-decoupled GNN training framework, which decouples different training tasks to manage their associated intermediate data independently and release it as early as possible to reduce memory usage. By integrating these techniques, we propose a multi-GPU GNN training system, NeutronTask. Experimental results on a 4×A5000 GPU server show that NeutronTask effectively supports billion-scale full-graph GNN training. For small graphs where the training data fits into the GPUs, NeutronTask achieves 1.27× - 5.47× speedup compared to state-of-the-art GNN systems including NeutronStar and Sancus.

Accurately predicting the trajectories of surrounding agents is critically important for the safe operation of autonomous vehicles. However, complex environments not only increase the computational cost but also introduce multimodality in the agent’s future motions, presenting significant challenges for motion prediction. Therefore, we propose FSGA, a trajectory prediction framework that constructs a relationship graph between agents and lanes in future spacetime and utilizes a graph attention mechanism to analyze the interactions between these future spacetime agents and lanes. We employ a vectorized representation method to model the traffic scene as a graph. Through this vectorized scene representation, FSGA extracts spatio-temporal features from agents’ historical trajectories, including interactions between agents, and between agents and lanes. To reduce computational costs, a lane scoring mechanism is employed, allowing the model to focus only on the lanes that have a significant impact on the current agent’s motion. To address complexity and multimodality of future trajectories, we introduce a branching tree goal prediction structure that models multiple possible motion trends by representing waypoints and endpoints at different temporal stages. Based on these predicted goals, we construct a relationship graph that captures the associations between agents and lanes in future spacetime and obtains the interactions between agents and lanes, improving trajectory prediction accuracy. Experiments show that FSGA achieves state-of-the-art performance on Argoverse1.

Multi-lane formation control can significantly enhance the efficiency of traffic systems in multi-lane scenarios. However, existing multi-lane formation control methods are mostly developed for fully Connected and Automated Vehicle (CAV) environments and lack a mechanism for multi-lane formation control in mixed traffic. This paper proposes a CAV formation grouping method and an interaction mechanism of CAVs and Human-driven Vehicles (HDVs), which is suitable for different traffic volumes and penetration rates. Depending on the positional relationship between the CAV formation and HDVs, it flexibly chooses between conventional formation control methods or improved mixed traffic formation control methods. By conducting simulations at various input flow volumes and penetration rates, the applicability of this method to different conditions is verified. The paper also explores the extent to which multi-lane formation control methods can improve traffic efficiency and their relationship with traffic volume and penetration rate.

In this article, we present the first experimental results of data-driven predictive control for connected and autonomous vehicles (CAVs) in dissipating traffic waves. In particular, we consider a recent strategy of Data-EnablEd Predictive Leading Cruise Control (DeeP-LCC), which bypasses the need of identifying the driving behaviors of surrounding vehicles and directly relies on measurable traffic data to achieve safe and optimal CAV control in mixed traffic. We present the implementation details ofDeeP-LCC, including data collection, equilibrium estimation, and control execution. Based on a miniature experiment platform, we reproduce the phenomenon of stop-and-go waves in two typical traffic scenarios: 1) open straight-road scenario under external disturbances and 2) closed ring-road scenario with no bottlenecks. Our experiments clearly demonstrate thatDeeP-LCCenables one or a few CAVs to dissipate the traffic waves in both traffic scenarios. These experimental findings validate the great potential ofDeeP-LCCin smoothing practical traffic flow in the presence of noisy data, uncertain low-level vehicle dynamics, and communication and computation delays. The code and videos of our experimental results are available athttps://github.com/soc-ucsd/DeeP-LCC.

Graph Neural Networks (GNNs) have shown exceptional performance across a wide range of applications. Current frameworks leverage CPU-GPU heterogeneous environments for GNN model training, incorporating mini-batch and sampling techniques to mitigate GPU memory constraints. In such settings, sample-based GNN training can be divided into three phases: sampling, gathering, and training. Existing GNN systems deploy various task orchestration methods to execute each phase on either the CPU or GPU. However, through comprehensive experimentation and analysis, we observe that these task orchestration approaches do not optimally exploit the available heterogeneous resources, hindered by either inefficient CPU processing or GPU resource bottlenecks. In this paper, we propose NeutronOrch, a system for sample-based GNN training that ensures balanced utilization of the CPU and GPU. NeutronOrch decouples the training process by layer and pushes down the training task of the bottom layer to the CPU. This significantly reduces the computational load and memory footprint of GPU training. To avoid inefficient CPU processing, NeutronOrch only offloads the training of frequently accessed vertices to the CPU and lets GPU reuse their embeddings with bounded staleness. Furthermore, NeutronOrch provides a fine-grained pipeline design for the layer-based task orchestrating method. The experimental results show that compared with the state-of-the-art GNN systems, NeutronOrch can achieve up to 11.51× performance speedup.

Graph neural networks (GNNs) have emerged as a promising direction. Training large-scale graphs that relies on distributed computing power poses new challenges. Existing distributed GNN systems leverage data parallelism by partitioning the input graph and distributing it to multiple workers. However, due to the irregular nature of the graph structure, existing distributed approaches suffer from unbalanced workloads and high overhead in managing cross-worker vertex dependencies. In this paper, we leverage tensor parallelism for distributed GNN training. GNN tensor parallelism eliminates cross-worker vertex dependencies by partitioning features instead of graph structures. Different workers are assigned training tasks on different feature slices with the same dimensional size, leading to a complete load balance. We achieve efficient GNN tensor parallelism through two critical functions. Firstly, we employ a generalized decoupled training framework to decouple NN operations from graph aggregation operations, significantly reducing the communication overhead caused by NN operations which must be computed using complete features. Secondly, we employ a memory-efficient task scheduling strategy to support the training of large graphs exceeding single GPU memory, while further improving performance by overlapping communication and computation. By integrating the above techniques, we propose a distributed GNN training system NeutronTP. Our experimental results on a 16-node Aliyun cluster demonstrate that NeutronTP achieves 1.29×-8.72× speedup over state-of-the-art GNN systems including DistDGL, NeutronStar, and Sancus.

Many Graph Neural Network (GNN) training systems have emerged recently to support efficient GNN training. Since GNNs embody complex data dependencies between training samples, the training of GNNs should address distinct challenges different from DNN training in data management, such as data partitioning, batch preparation for mini-batch training, and data transferring between CPUs and GPUs. These factors, which take up a large proportion of training time, make data management in GNN training more significant. This paper reviews GNN training from a data management perspective and provides a comprehensive analysis and evaluation of the representative approaches. We conduct extensive experiments on various benchmark datasets and show many interesting and valuable results. We also provide some practical tips learned from these experiments, which are helpful for designing GNN training systems in the future.

Car-following behavior is one of the most common driving behaviors. To reduce the impact of driver reaction delay and accumulated errors in predicting long sequences on the accuracy of speed prediction, we propose a deep learning car-following model based on a sequence-to-sequence (seq2seq) architecture with an attention mechanism. Firstly, we analyze the characteristics of driver reaction delay during the car-following process and design an attention mechanism to learn the probability distribution of driver reaction delay. This allows the model to consider more environmental information at the moment when the driver actually makes a decision, rather than just the current environmental information, during the prediction process. By utilizing the seq2seq architecture to model car-following behavior, the model focuses more on reducing the impact of accumulated errors during the training process. Additionally, we propose a temporal consistency constraint loss to improve the robustness and stability of the car-following model’s training method and enhance the prediction results. Finally, we use Gated Recurrent Units (GRU) and the sequence-to-sequence model (seq2seq) as baselines, and simulation results demonstrate that our model achieves more accurate car-following behavior prediction.

In industrial practice, GUI (Graphic User Interface) testing of mobile apps still inevitably relies on huge manual efforts. The major efforts are those on understanding the GUIs, so that testing scripts can be written accordingly. Quality assurance could therefore be very labor-intensive, especially for modern commercial mobile apps, where one may include tremendous, diverse, and complex GUIs, e.g., those for placing orders of different commercial items. To reduce such human efforts, we propose Appaction, a learning-based automatic GUI interaction approach we developed for Meituan, one of the largest E-commerce providers with over 600 million users. Appaction can automatically analyze the target GUI and understand what each input of the GUI is about, so that corresponding valid inputs can be entered accordingly. To this end, Appaction adopts a multi-modal model to learn from human experiences in perceiving a GUI. This allows it to infer corresponding valid input events that can properly interact with the GUI. In this way, the target app can be effectively exercised. We present our experiences in Meituan on applying Appaction to popular commercial apps. We demonstrate the effectiveness of Appaction in GUI analysis, and it can perform correct interactions for numerous form pages.

Existing Graph Neural Network (GNN) training frameworks have been designed to help developers easily create performant GNN implementations. However, most existing GNN frameworks assume that the input graphs are static, but ignore that most real-world graphs are constantly evolving. Though many dynamic GNN models have emerged to learn from evolving graphs, the training process of these dynamic GNNs is dramatically different from traditional GNNs in that it captures both the spatial and temporal dependencies of graph updates. This poses new challenges for designing dynamic GNN training frameworks. First, the traditional batched training method fails to capture real-time structural evolution information. Second, the time-dependent nature makes parallel training hard to design. Third, it lacks system supports for users to efficiently implement dynamic GNNs. In this paper, we present NeutronStream, a framework for training dynamic GNN models. NeutronStream abstracts the input dynamic graph into a chronologically updated stream of events and processes the stream with an optimized sliding window to incrementally capture the spatial-temporal dependencies of events. Furthermore, NeutronStream provides a parallel execution engine to tackle the sequential event processing challenge to achieve high performance. NeutronStream also integrates a built-in graph storage structure that supports dynamic updates and provides a set of easy-to-use APIs that allow users to express their dynamic GNNs. Our experimental results demonstrate that, compared to state-of-the-art dynamic GNN implementations, NeutronStream achieves speedups ranging from 1.48X to 5.87X and an average accuracy improvement of 3.97%.

The cooperation between connected and automated vehicles (CAVs) has emerged as a promising way to improve traffic efficiency and safety for ramp merging on highways. Existing research mostly focused on the collaboration of individual CAVs, while the cooperative merging strategy in mixed traffic considering vehicle platoons has been less explored. To address the above problem, this study aims to build connections between sequence scheduling and motion planning in mixed traffic, where individual CAVs, CAVs platoons, and mixed platoons coexist. First, optimal control strategies are presented for vehicles with flexible merging points based on Pontryagin’s minimum principle (PMP), and the final vehicle states in variable conditions are summarized in a phase diagram. Subsequently, the optimal final-state phase diagram is introduced into the passing sequence tree search process, which is designed for different vehicle groups in mixed traffic. Heuristic pruning rules are added to the depth-first search strategy to facilitate finding the optimal solution. Finally, an event-triggered receding horizon optimization algorithm is developed for continuous implementation. The numerical simulations are conducted in multiple traffic volumes, and the simulation results reveal that our proposed algorithm significantly improves the overall traffic efficiency and reduces vehicle-passing delays compared with the traditional FIFO-based cooperation method.

GNN's training needs to resolve issues of vertex dependencies, i.e., each vertex representation's update depends on its neighbors. Existing distributed GNN systems adopt either a dependencies-cached approach or a dependencies-communicated approach. Having made intensive experiments and analysis, we find that a decision to choose one or the other approach for the best performance is determined by a set of factors, including graph inputs, model configurations, and an underlying computing cluster environment. If various GNN trainings are supported solely by one approach, the performance results are often suboptimal. We study related factors for each GNN training before its execution to choose the best-fit approach accordingly. We propose a hybrid dependency-handling approach that adaptively takes the merits of the two approaches at runtime. Based on the hybrid approach, we further develop a distributed GNN training system called NeutronStar, which makes high performance GNN trainings in an automatic way. NeutronStar is also empowered by effective optimizations in CPU-GPU computation and data processing. Our experimental results on 16-node Aliyun cluster demonstrate that NeutronStar achieves 1.81X-14.25X speedup over existing GNN systems including DistDGL and ROC.

Connected and automated vehicles have shown great potential in improving traffic mobility and reducing emissions, especially at unsignalized intersections. Previous research has shown that vehicle passing order is the key influencing factor in improving intersection traffic mobility. In this paper, we propose a graph-based cooperation method to formalize the conflict-free scheduling problem at an unsignalized intersection. Based on graphical analysis, a vehicle’s trajectory conflict relationship is modeled as a conflict directed graph and a coexisting undirected graph. Then, two graph-based methods are proposed to find the vehicle passing order. The first is an improved depth-first spanning tree algorithm, which aims to find the local optimal passing order vehicle by vehicle. The other novel method is a minimum clique cover algorithm, which identifies the global optimal solution. Finally, a distributed control framework and communication topology are presented to realize the conflict-free cooperation of vehicles. Extensive numerical simulations are conducted for various numbers of vehicles and traffic volumes, and the simulation results prove the effectiveness of the proposed algorithms.

Connected and autonomous vehicles (CAVs) have great potential to improve road transportation systems. Most existing strategies for CAVs’ longitudinal control focus on downstream traffic conditions, but neglect the impact of CAVs’ behaviors on upstream traffic flow. In this paper, we introduce a notion of Leading Cruise Control (LCC), in which the CAV maintains car-following operations adapting to the states of its preceding vehicles, and also aims to lead the motion of its following vehicles. Specifically, by controlling the CAV, LCC aims to attenuate downstream traffic perturbations and smooth upstream traffic flow actively. We first present the dynamical modeling of LCC, with a focus on three fundamental scenarios: car-following, free-driving, and Connected Cruise Control. Then, the analysis of controllability, observability, and head-to-tail string stability reveals the feasibility and potential of LCC in improving mixed traffic flow performance. Extensive numerical studies validate that the capability of CAVs in dissipating traffic perturbations is further strengthened when incorporating the information of the vehicles behind into the CAVs’ control.

This paper presents a general design of an autonomous driving platform. Firstly, we design a physic architecture and three-bus power system to guarantee safety and power redundancy. The logical architecture of the platform is designed in five sub-functions. In most of the cases, planning and control function is the critical and universal function, where we use curvature and environmental perception information to calculate vehicle speed. We also apply feedforward-feedback controller with preview point method to calculate steering angle. In the end, we conduct field test and road test to verify the functionality of the platform. Different navigation methods are also compared in the tests.