Mobile agents have made progress toward reliable smartphone automation, yet performance in complex applications remains limited by incomplete knowledge and weak generalization to unseen environments. We introduce a curiosity driven knowledge retrieval framework that formalizes uncertainty during execution as a curiosity score. When this score exceeds a threshold, the system retrieves external information from documentation, code repositories, and historical trajectories. Retrieved content is organized into structured AppCards, which encode functional semantics, parameter conventions, interface mappings, and interaction patterns. During execution, an enhanced agent selectively integrates relevant AppCards into its reasoning process, thereby compensating for knowledge blind spots and improving planning reliability. Evaluation on the AndroidWorld benchmark shows consistent improvements across backbones, with an average gain of six percentage points and a new state of the art success rate of 88.8% when combined with GPT-5. Analysis indicates that AppCards are particularly effective for multi step and cross application tasks, while improvements depend on the backbone model. Case studies further confirm that AppCards reduce ambiguity, shorten exploration, and support stable execution trajectories. Task trajectories are publicly available at https://lisalsj.github.io/Droidrun-appcard/.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Modern Large Language Model (LLM) training clusters increasingly mix heterogeneous GPUs, diverse intra-node fabrics, and inter-node interconnects, combined with varied parallelism strategies. Exploring this massive design space, further amplified by heterogeneity, through real deployments is prohibitively slow and costly. Existing simulators, which are primarily designed and tuned for homogeneous clusters, either trade fidelity for speed or require heavyweight workflows with non-negligible overhead. We propose HeteroSim, a high-fidelity simulation framework for heterogeneous LLM training systems. It introduces: (i) a LLM training workload compiler that captures realistic training graphs, microbatching schedules, and compute-communication overlap; (ii) a heterogeneity-aware computation planner using roofline-style scaling across GPU generations; (iii) a collective communication planner that reproduces NCCL-like behaviors with per-link models, message channelization, and configurable routing. Across a wide range of heterogeneity levels, experimental results show that HeteroSim achieves near-real simulation accuracy while keeping low overhead.
Cross-view geo-localization (CVGL) establishes correspondences between ground-level and satellite images of the same geographic location, serving as a fundamental technology for smart city applications, including autonomous navigation, urban planning, and location-based services. Current CVGL approaches fall into two categories: feature-based methods achieve superior performance through 2D representation learning but lack interpretability. Spatial-based methods provide geometric understanding and interpretable matching but suffer from limited spatial modeling and weak cross-view alignment, leading to lower performance. We reformulate CVGL from a spatial perspective and propose an auxiliary task-enhanced network. The network captures spatial semantics and provides explicit alignment processes with visualizable results. We introduce an auxiliary spatial semantic alignment (SSA) task that learns spatial structure via vision foundation models (VFM) and BEV transformation to enhance the primary CVGL task. The primary task captures visual semantics, including texture and appearance. Within this unified framework with shared encoders, the primary task enriches the learned embeddings by fusing spatial structure with visual semantics, yielding spatially complete representations. Extensive experiments on three standard CVGL benchmarks demonstrate that our method significantly surpasses previous spatial-based approaches while maintaining competitive performance with state-of-the-art (SOTA) feature-based methods, achieving 98.48% R@1 on CVUSA and 71.05% R@1 on CVACT\_test. We provide comprehensive analyses through pixel-level activation maps and feature-space UMAP visualizations to validate both effectiveness and interpretability.
Low Earth Orbit (LEO) satellite ISPs promise universal Internet connectivity, yet their interaction with content delivery remains poorly understood. We present the first comprehensive measurement study decomposing Starlink's web content delivery performance decomposed across Point of Presence (PoP), DNS, and CDN layers. To quantify how satellite architecture disrupts terrestrial CDN assumptions, we conduct a measurement study spanning two years. We identify three distinct performance regimes based on infrastructure density. Regions with local content-rich PoPs achieve near-terrestrial latencies with the satellite segment dominating 80–90% of RTT. Infrastructure-sparse regions suffer cascading penalties: remote PoPs force distant resolver selection, which triggers CDN mislocalization, pushing latencies beyond 200 ms. Dense-infrastructure regions show minimal sensitivity to PoP changes. Leveraging Starlink's infrastructure expansion in early 2025 as a natural experiment, we demonstrate that relocating PoPs closer to user location reduces median page-fetch times by 60%. Our findings reveal that infrastructure proximity, not satellite coverage, influences web performance, requiring fundamental changes to CDN mapping and DNS resolution for satellite ISPs.
The sparse mixture-of-experts (MoE) architecture is a promising backbone of foundation models for a wide range of applications in edge. However, deploying them locally presents a significant challenge to memory-constrained GPUs. Previous techniques utilize CPUs for expert offloading, which suffer from inaccurate expert prefetching and on-demand loading latency. To address these challenges, we propose self-speculative MoE ( SS-MoE ), an algorithm-system co-design framework that facilitates inference under limited GPU memory. Our insight is that only a subset of routed experts, i.e., draft model, can still tackle easy tasks and generate draft tokens. Second, we deem GPU memory as the experts cache, and on-demand update it to mitigate IO overhead. Draft tokens from fewer routed experts are generated quickly, and these experts are then routed for verification. Additionally, we design a confidence-based policy to adaptively accept or verify draft tokens, which selectively decreases or increases the number of verification tokens of speculative decoding and achieves acceleration. Notably, under conservative verification, our approach preserves model accuracy and surpasses the decoding speed of the 4-bit quantized counterpart model. Under adaptive verification, our method significantly enhances decoding speed by 3.72x over state-of-the-art methods while maintaining nearly lossless accuracy.
WebAssembly (Wasm) is a web standard that defines a binary instruction format serving as a portable compilation target for high-level programming languages. As a cornerstone design feature, code porting empowers developers to migrate code across platforms without extensive redevelopment. However, we observe that certain native code exhibits functional discrepancies compared to its Wasm counterpart. These discrepancies impede migration by causing functional alterations, generating misleading messages, and forcing platform-specific workarounds. In this paper, we present an in-depth analysis of these Native-to-Wasm discrepancies, demonstrating that they often stem from complex underlying causes, leading to unexpected outcomes in real-world applications. We characterize the root causes and practical implications of these discrepancies, providing critical insights into WebAssembly code porting challenges. To facilitate this large-scale analysis, we develop WasmDiff, an automated differential fuzzing framework that detects Native-to-Wasm discrepancies by comparing the execution of identical code across platforms to identify input-triggered behavioral differences. We identify 14,053 discrepancies across 64,001 test samples from the basic benchmark and 3,102 discrepancies within eight real-world projects. Our findings reveal that discrepancies are prevalent, input-dependent and rooted in diverse execution stacks, underscoring the need for greater attention to code porting issues in the WebAssembly ecosystem.
Generative modeling offers a promising solution to data scarcity and privacy challenges in time series analysis. However, the structural complexity of time series, characterized by multi-scale temporal patterns and heterogeneous components, remains insufficiently addressed. In this work, we propose a structure-disentangled multiscale generation framework for time series. Our approach encodes sequences into discrete tokens at multiple temporal resolutions and performs autoregressive generation in a coarse-to-fine manner, thereby preserving hierarchical dependencies. To tackle structural heterogeneity, we introduce a dual-path VQ-VAE that disentangles trend and seasonal components, enabling the learning of semantically consistent latent representations. Additionally, we present a guidance-based reconstruction strategy, where coarse seasonal signals are utilized as priors to guide the reconstruction of fine-grained seasonal patterns. Experiments on six datasets show that our approach produces higher-quality time series than existing methods. Notably, our model achieves strong performance with a significantly reduced parameter count and exhibits superior capability in generating high-quality long-term sequences. Our implementation is available at https://github.com/decisionintelligence/TimeMAR.
Trace analysis is essential for understanding system behaviors, detecting anomalies, and diagnosing faults in complex microservice-based web applications. Existing trace analysis approaches face several challenges in industrial microservice-based systems, including high manual overhead, limited functionality, unfriendly interaction mechanisms, and difficulties in deployment and integration. The strong capabilities of large language models (LLMs) in natural language understanding, reasoning, and multi-task generalization provide new opportunities for a more intelligent and flexible trace analysis approach. However, the trace analysis capabilities of LLMs remain underexplored and underdeveloped. To bridge this gap, we conduct the first comprehensive evaluation on the trace analysis capabilities of LLMs. In particular, we construct the first instruction&response benchmark dataset for trace analysis, named TraceBench. It involves a wide range of trace analysis tasks, allowing us to systematically evaluate the capabilities of LLMs in this area. Experimental results show that LLMs have potential in handling trace analysis tasks, but there leaves room for improvement. To this end, we propose TraceLLM, an approach that significantly enhances the capabilities of LLMs via fine-tuning, outperforming the open-source LLMs by 34.77% on average in terms of accuracy, and outperforming the closed-source model by 21.66% in the best case. The generalization and robustness of TraceLLM are also confirmed in our experiments. To the best of our knowledge, TraceLLM is the first LLM which is specialized for handling various types of trace analysis tasks. This work provides a foundation for future research to further explore the trace analysis capabilities of LLMs.
The web-scale surge of multimodal content, including short-video feeds and autonomous sensing streams, has made web-native lossless compression a prerequisite for delivery and storage across browsers and edge–cloud pipelines. However, existing methods often fail to adapt to shifting distributions across different batches and struggle to balance computational resources in the face of large conditional entropy disparities among diverse modalities. To address these limitations, we propose MoE-LC, a new mixture-of-experts framework for multi-modal lossless compression that dynamically accommodates heterogeneous data distributions and varying complexity levels. First, the Batch-Adaptive Experts (BAE) module introduces batch-specific parameters with a residual gating mechanism, ensuring stable modeling under non-stationary distributions. Second, the Entropy-Aware Multi-Expert Selection (MES) strategy adaptively allocates the number of experts according to the data's estimated compression difficulty (entropy), thereby improving resource utilization and computational efficiency. Finally, the Precision-Aware Expert Routing (PER) component applies high-precision computation solely to the most critical experts, significantly reducing overhead without sacrificing compression accuracy. Experimental results across multiple real-world datasets demonstrate that MoE-LC achieves 5.33%--70.89% improvements in compression ratio and 37.25%--1532.41% gains in throughput compared to advanced baselines, offering a scalable solution for real-time, large-scale multi-modal data compression. Our code is available at https://github.com/Magie0/MoE_LC.
Large language models (LLMs) are increasingly powering web-based applications, whose effectiveness relies on fine-tuning with large-scale instruction data. However, such data often contains valuable or sensitive information that limits its public sharing among business organizations. Federated learning (FL) enables collaborative fine-tuning of LLMs without accessing raw data. Existing approaches to federated LLM fine-tuning usually adopt a uniform model architecture, making it challenging to fit highly heterogeneous client-side data in varying domains and tasks, e.g., hospitals and financial institutions conducting federated fine-tuning may require different LLM architectures due to the distinct nature of their domains and tasks. To address this, we propose FedAMoLE, a lightweight personalized FL framework that enables data-driven heterogeneous model architectures. It features a heterogeneous mixture of low-rank adaptation (LoRA) experts module to aggregate architecturally heterogeneous models and a reverse selection-based expert assignment strategy to tailor model architectures for each client based on data distributions. Experiments across seven scenarios demonstrate that FedAMoLE improves client-side performance by an average of 5.97% over existing approaches while maintaining practical memory, communication, and computation overhead.
The current paradigm of training large language models (LLMs) on public available Web data is becoming unsustainable as high-quality data sources in specialized domains near exhaustion. Federated Learning (FL) emerges as a practical solution for the next generation of AI on a decentralized Web, enabling privacy-preserving collaborative fine-tuning on decentralized private data. While Low-Rank Adaptation (LoRA) is standard for efficient fine-tuning, its federated application faces a critical bottleneck: communication overhead under heterogeneous network conditions. Structural redundancy in LoRA parameters increases communication costs and causes aggregation conflicts. To address this, we propose FedSRD, a Sparsify-Reconstruct-Decompose framework for communication-efficient federated LLM fine-tuning. We introduce importance-aware sparsification to reduce the upload parameter count while preserving the structural integrity of LoRA updates. The server aggregates updates in full-rank space to mitigate conflicts, then decomposes the global update into a sparse low-rank format for broadcast, ensuring a symmetrically efficient cycle. We also propose an efficient variant, FedSRD-e, to reduce computational overhead. Experiments on 10 benchmarks show our framework significantly reduces communication costs by up to 90% while improving performance on heterogeneous client data.
Federated Graph Learning (FGL) has emerged as an efficient paradigm to address the pronounced data heterogeneity common in real-world decentralized graph datasets, attracting significant interest from both academia and industry. Existing FGL methods often degrade the global model's performance due to a fundamental mismatch between the model's fixed-geometry embedding space and the diverse geometric structures of client data. To address this issue, we propose a novel framework for Personalized Federated Riemannian Graph Learning, namely FedRGL. FedRGL introduces a personalized mixed-curvature product space for each heterogeneous client, mapping local graph data into a tailored geometric space composed of Euclidean, hyperbolic, and spherical manifolds. Furthermore, it leverages a Ricci-Gated Graph Convolutional Network to dynamically adapt its message-passing mechanism to the local topology of each graph. Extensive experiments across diverse geometric heterogeneity settings demonstrate that FedRGL significantly outperforms state-of-the-art FGL methods in terms of model accuracy and generalization.
The rapid development of the Internet of Things (IoT) applications necessitates resource-efficient computing paradigms that can unify heterogeneous sensing modalities. Spiking Neural Networks (SNNs) meet this need with their event-driven and energy-efficient processing nature. However, deploying SNNs on mobile and embedded platforms is hindered by strict and fluctuating memory budgets. While prior work explores lightweight model design and system-level memory management, these methods either sacrifice accuracy or incur high runtime overhead due to timestep-dependent dynamics. To tackle these challenges, we propose a memory-adaptive framework MASI that enables efficient on-device SNN inference by combining (1) a fine-grained memory-adaptive layer slicing strategy, (2) a timestep-agnostic scheduler that maximizes memory utilization with minimal fragmentation, and (3) a timestep-aware early-exit mechanism that reduces redundant calculations. Evaluated on diverse workloads and edge devices, MASI can dynamically adapt to runtime memory availability, approximately reducing memory usage by 20.67% and inference latency by 58.53% on average with negligible accuracy loss compared to other feasible on-device implementations under memory constraints.
Virtual private clouds (VPCs) play a critical role in providing secure and isolated network environments for web services. However, with the growing number and size of VPCs, efficiently delivering control messages from the control plane to the data plane has become a major concern for cloud vendors. Existing end-to-end transmission solutions (e.g., RPC) will result in substantial overhead in the control plane, while message-oriented middleware-based solutions (e.g., message queue) will lead to high data plane overhead. To address this issue, we design Meteor, a high-performance control message delivery system for large-scale clouds. Specifically, Meteor combines an RPC path with a message queue (MQ) path and employs an auto dual-path switching mechanism to minimize the message delivery latency. Additionally, we propose a VPC-based message delivery and filtering scheme for the MQ path to reduce data plane overhead. We also design a delivery robustness guarantee mechanism to ensure the reachability and consistency of control messages. Meteor has been thoroughly tested with up to 100k container instances. Evaluation results show that Meteor decreases the message delivery latency by 48.8% and reduces the overhead by about 50% in real-world scenarios, compared with state-of-the-art solutions.
Traffic prediction serves as a cornerstone for systems and network services such as the Web of Vehicles (WoV), online navigation, and smart city applications. Despite the proliferation of model architectures in recent years, existing approaches often suffer from highly customized structures and weak transferability, making it difficult to cope with increasing task heterogeneity and modeling complexity. To address these challenges, we propose ST-LEGO, a modular assembly framework driven by large language models (LLMs) that supports flexible structural composition and automated code generation. ST-LEGO employs a multi-agent collaborative system comprising a Prompt Agent, Assemble Agent, and Code Agent, which are responsible for understanding task requirements, dynamically assembling structural modules, and automatically generating executable PyTorch code. By introducing a standardized module library and an intermediate structural description language (DSL), the framework enables controllable generation, reusable composition, and cross-task generalization of model architectures. Empirical results on multiple real-world traffic datasets demonstrate that models generated by ST-LEGO achieve superior accuracy, structural diversity, and convergence compared to a wide range of manually designed baselines. These results highlight the unique potential and scalability of LLMs as structural architects for traffic prediction, offering a new paradigm for integrating language models into web-interactive intelligent transportation systems.
Drones are increasingly used for on-device video analytics, yet their limited resources can only support lightweight DNNs. To enhance on-device models, continuous learning has been proposed that adapts DNNs to real-time videos by model reuse or retraining. However, existing methods struggle with low accuracy due to drastic drift in agile drones' videos, which leads to outdated retraining, false alarm in drift detection, and intricate performance kinship within the historical model zoo. In this paper, we propose SkyCL, a swift continuous learning system to enhance multi-drone video analytics under drastic drift. SkyCL achieves swift on-device adaptation by dual-level drift detection and distributed retraining scheduling across multiple edge devices. Moreover, SkyCL leverages model kinship within the model zoo to discover a hidden path of models' improvement for potential optimal reuse. We evaluate SkyCL in both simulated and real drone systems, with results showing a 14.39% accuracy gain and a 2.41× speedup over baselines.
Large-scale vision-language models enable powerful cross-modal understanding and generation, driving rapidly growing demand for online inference services. However, cloud-centric serving often suffers from high latency, rising costs, and network dependency, while purely on-device deployment is constrained by limited memory and reduced accuracy on complex tasks. To address this accuracy–latency–cost trilemma, we propose ShiftVL, a task-aware end–cloud serving framework that shifts suitable execution to the end device with a cloud fallback. ShiftVL serves high-frequency requests on an end-side small VLM enhanced with ViTexLoRA, a modality-disentangled parameter-efficient tuning method that preserves cross-modal alignment, while routing low-frequency or complex requests to a cloud-hosted large VLM for higher accuracy. Under tight device budgets, ShiftVL employs a predictive adapter scheduler that combines LRU-style caching with imitation learning to pre-load task-specific adapters. Experiments with InternVL models show that ShiftVL reduces cloud cost by up to 76.3% and latency by up to 42.9% while maintaining high multi-task accuracy, demonstrating its practicality for real-world vision-language model serving.
As microservice architecture is evolving toward Logical Data Center (LDC), accurate forecasting of the microservices indicators can support reasonable resource allocation, thereby ensuring the availability and reliability of cloud service. From a macro perspective, due to the architecture hierarchy, microservices exhibit: 1) collaborative relationships derived from shared functionalities, 2) backup relationships between replicas, and 3) dynamic correlation driven by cooperation. From a micro perspective, there exist causal relationships among indicators within a microservice. That is, workload will first impact system consumption, such as CPU and memory usage, then affect service quality like system latency. Based on these insights, we propose MaMiClif, a macro-micro collaborative learning framework for LDC microservice indicators forecasting. MaMiClif constructs Macro Graph and Micro Matrix to model the microservices dependencies and the causality of indicators. To learn fine-grained indicator dependencies, Indicator-Centric Embedding is leveraged to generate representations for indicator series. We use Heterogeneous Graph Convolution to update workload representations based on the Macro Graph, and adopt Causal Sparse Self-attention to integrate causal strength into the self-attention calculation, enabling a comprehensive exploration of dependencies among indicators. Experiments on two datasets, including LDC_MS, which was collected from the LDC system of Ant Group, demonstrate the effectiveness of MaMiClif.
Pre-trained visual models have been deployed on various edge devices to facilitate a broad range of downstream tasks through few-shot learning (FSL), in particular when downstream data or on-device resources are limited. However, FSL often suffers from poor performance due to its inability to adapt to the characteristics of the deployment environments, while backbone fine-tuning prior to model deployment is typically infeasible because of the unavailability of environment-specific samples. To tackle this challenge, this paper presents Glasses, a lightweight fine-tuning scheme that can adapt ViT-based model backbones to deployment environments rapidly through device-cloud collaboration, helping the model achieve better FSL performance on the device. Glasses leverages the computational power and sample resources in the cloud to produce model updates for rapid model adaptation based on only one environment image without needing a label. Experiments with five models on two datasets demonstrate that Glasses can adapt a model rapidly and outperform the original backbone by 5.54%–22.56% in the 1-shot setting and by 2.50%–10.28% in the 5-shot setting. The source code is available at https://github.com/CGCL-codes/Glasses.
Digital maps are crucial for web-based location services. The continuous collection of vehicle trajectories has made trajectory data a vital source for map inference. State-of-the-art (SOTA) methods formulate trajectory-based map inference as an image processing task: they first rasterize trajectories into density-based images, then extract keypoints and infer their connections to construct the road map. Despite performing well on standard road structures, these methods still suffer from low topological accuracy in complex scenarios because i) The discrete rasterized representation struggles to capture road adjacency in multi-level road structures. ii) The limited contextual awareness of keypoint-based inferring strategy leads to connectivity misjudgment in dense road areas. To solve these limitations, we propose D2Map, a Dual-view Map Inference Framework via Primal-Dual Graphs Co-generation. To precisely encode road adjacency, we introduce the serialized trajectory view as a complement to the rasterized view to reflect traversable relationships between roads, and devise a strategy-adaptive fusion module that dynamically selects and executes the optimal fusion operator to integrate dual view representations, yielding map element embeddings. To eliminate connectivity errors, we extend road map modeling from a keypoint-centric primal graph to primal and dual graphs. In the dual graph, roads are explicitly modeled as nodes, enabling context-aware topology inference. A co-generation strategy is then employed to jointly infer both graphs while maintaining their geometric consistency. Extensive experiments on two real-world datasets demonstrate the superiority of D2Map, which outperforms SOTA baselines by 11.44% in the TOPO metric.