Searcharxiv⌕ Search

arXiv subjects

Leyang Xue

Publications and source records attributed to Leyang Xue.

At least 19 recordsLinked to original sources
arXiv2026

Distinct routes to phase transitions in spatial activation systems

Threshold-driven activation governs a wide range of collective phenomena, yet the microscopic origins of its phase transitions in spatial systems remain unresolved. Here, we show that spatial activation systems undergo multiple distinct routes to phase transitions, controlled by a single parameter---the interaction range. We uncover a unified phase diagram featuring continuous, first-order, and mixed-order transitions, and demonstrate that the two abrupt transitions arise from fundamentally different mechanisms: nucleation-driven front propagation and critical branching. These routes exhibit distinct dynamical scaling, establishing a direct link between microscopic activation dynamics and macroscopic critical behavior. We further identify a metastable phase in which global activation cannot be achieved by random activation alone, but can be triggered by localized seeds. In this regime, the critical activation nucleus remains finite and independent of system size, implying that arbitrarily large systems can remain stable under random perturbations yet highly vulnerable to localized triggers. The onset of this phase is abrupt, revealing an extreme sensitivity of collective dynamics to small parameter changes. These results establish a mechanistic framework for phase transitions in spatial activation systems and reveal how microscopic perturbations can trigger macroscopic cascades.

physics.soc-ph↗
arXiv2026

BatchGen: An Architecture for Scalable and Efficient Batch Inference

Batch inference has become a central mode of AI computation, yet existing inference engines still rely on execution models designed for interactive serving. When scaled to millions of sequences, batch workloads reveal two fundamental requirements: the ability to handle extreme inter- and intra-sequence load variation that emerges only at runtime, and the ability to sustain high utilization across large fleets of GPUs. Existing systems fail to meet these requirements, losing substantial fractions of achievable throughput. We introduce a new architectural foundation for batch inference: the sequence coroutine compute model, which represents each sequence as a fine-grained, event-driven coroutine. This model exposes expressive primitives that allow the runtime to reorganize work dynamically, enabling larger expert-level batches, mitigating stragglers, reallocating work across devices, and maintaining utilization even on cost-effective or memory-constrained GPUs. Building on this abstraction, we implement BatchGen, a production-ready system that uses the coroutine model at cluster scale. On a 128-GPU cluster, BatchGen reduces batch completion time by up to $2.3\times$, and on memory-constrained accelerators it outperforms the strongest offloading baseline by up to $9.6\times$. We will open-source BatchGen at https://github.com/batchgen-project/batchgen

cs.DC↗
arXiv2026

TUBO: A Tailored ML Framework for Reliable Network Traffic Forecasting

Traffic forecasting based network operation optimization and management offers enormous promise but also presents significant challenges from traffic forecasting perspective. While deep learning models have proven to be relatively more effective than traditional statistical methods for time series forecasting, their reliability is not satisfactory due to their inability to effectively handle unique characteristics of network traffic. In particular, the burst and complex traffic patterns makes the existing models less reliable, as each type of deep learning model has limited capability in capturing traffic patterns. To address this issue, we introduce TUBO, a novel machine learning framework custom designed for reliable network traffic forecasting. TUBO features two key components: burst processing for handling significant traffic fluctuations and model selection for adapting to varying traffic patterns using a pool of models. A standout feature of TUBO is its ability to provide deterministic predictions along with quantified uncertainty, which serves as a cue for identifying the most reliable forecasts. Evaluations on three real-world network demand matrix (DM) datasets (Abilene, GEANT, and CERNET) show that TUBO significantly outperforms existing methods on forecasting accuracy (by 4 times), and also achieves up to 94% accuracy in burst occurrence forecasting. Furthermore, we also consider traffic demand forecasting based proactive traffic engineering (TE) as a downstream use case. Our results show that compared to reactive approaches and proactive TE using the best existing DM forecasting methods, proactive TE powered by TUBO improves aggregated throughput by 9 times and 3 times, respectively.

cs.LG↗
arXiv2026

The mechanistic origin of branching-driven nucleation in abrupt phase transitions

Phase transitions are the macroscopic manifestation of microscopic processes that drive a system towards a new state. The detailed evolution of these processes, particularly in abrupt phase transitions, are currently not fully understood. Here, we introduce a theoretical framework based on internal node dependencies within a single-layer lattice. Crucially, we demonstrate that the fundamental mechanism underlying abrupt transitions is nucleation propagation preceded by a slow cascading process which scales with the range of dependencies. Our findings show that the synergy between these two distinct stages is essential for the occurrence of an abrupt transition. The first stage of a slow cascading mechanism was recently observed experimentally in superconducting layered materials, where heat acts as the dependency links, for the limit of infinite dependency range. Our model thus generalizes the framework to include finite dependency ranges, revealing previously unobserved mechanisms that could be experimentally verified through controlling the range of thermal diffusion in the material. As a universal mechanism, our model provides a robust method to test nucleation-controlled phase transitions in multiple systems, providing a path to discover and understand microscopic mechanisms in phase transitions.

physics.soc-ph↗
arXiv2025

Network localization governs social contagion dynamics with macro-level reinforcement

The spread of ideas, behaviors, and technologies generally depends on feedback mechanisms operating across multiple scales. Previous studies have extensively examined pairwise transmission and local reinforcement. However, the role of macro-level social influence -- where widespread adoption enhances further adoption -- remains understudied. Here, we focus on a contagion process that incorporates both pairwise interactions and macro-level reinforcement. We show that the contagion undergoes a shift from continuous to mixed-order transition as macro-level influence exceeds a reinforcement threshold. Simulations on various real-world networks indicate that network localization governs the contagion outcomes by determining the critical point and the reinforcement threshold. Building on this insight, we develop a structural metric linking network localization to contagion dynamics, revealing a key trade-off: networks that facilitate weak contagion tend to experience slower diffusion and lower adoption rates, while networks that suppress weak contagions enable faster and more widespread adoption. These findings challenge the conventional belief that stronger local connectivity uniformly promotes contagion.

physics.soc-ph↗
arXiv2025

On Harnessing Idle Compute at the Edge for Foundation Model Training

The foundation-model ecosystem remains highly centralized because training requires immense compute resources and is therefore largely limited to large cloud operators. Edge-assisted foundation model training that harnesses spare compute on edge devices offers a more democratized alternative. However, existing edge-training approaches fall short: they struggle to match cloud-training performance, scale to larger models, fit within device memory limits, or keep communication overhead manageable. They also do not handle device heterogeneity and churn satisfactorily. We introduce Cleave, built on a structural insight: each GEMM has an asymmetric I/O pattern -- its input matrices, sent over downlink, are much larger than the partial output blocks returned over uplink -- matching edge networks where downlink bandwidth exceeds uplink by 2--10x. Exploiting this alignment with a parameter-server-centric architecture, Cleave makes per-device communication \emph{decrease} as more devices join, rather than stay constant as in conventional TP. Decomposing training into independent sub-GEMM tasks yields one scheduling abstraction that unifies memory constraints, communication overhead, and fault tolerance under device churn. Our evaluation shows that Cleave achieves cloud-comparable GPU training performance and outperforms state-of-the-art edge-training methods by 4--10x in per-batch runtime at the same device counts. Beyond this shared operating range, Cleave scales to thousands of heterogeneous devices -- a regime where prior edge-training systems cannot operate -- and achieves at least 100x faster recovery from device failures.

cs.DC↗
arXiv2025

Towards Decentralized and Sustainable Foundation Model Training with the Edge

Foundation models are at the forefront of AI research, appealing for their ability to learn from vast datasets and cater to diverse tasks. Yet, their significant computational demands raise issues of environmental impact and the risk of centralized control in their development. We put forward a vision towards decentralized and sustainable foundation model training that leverages the collective compute of sparingly used connected edge AI devices. We present the rationale behind our vision, particularly in support of its sustainability benefit. We further outline a set of challenges that need to be addressed to turn this vision into reality.

cs.LG↗
arXiv2025

MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently, but it depends on heterogeneous compute and memory resources. These factors jointly affect system Cost, Accuracy, and Performance (CAP), making trade-offs inevitable. Existing benchmarks often fail to capture these trade-offs accurately, complicating practical deployment decisions. To address this, we introduce MoE-CAP, a benchmark specifically designed for MoE systems. Our analysis reveals that achieving an optimal balance across CAP is difficult with current hardware; MoE systems typically optimize two of the three dimensions at the expense of the third-a dynamic we term the MoE-CAP trade-off. To visualize this, we propose the CAP Radar Diagram. We further introduce sparsity-aware performance metrics-Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU)-to enable accurate performance benchmarking of MoE systems across diverse hardware platforms and deployment scenarios.

cs.LG↗
arXiv2025

HybridServe: Efficient Serving of Large AI Models with Confidence-Based Cascade Routing

Giant Deep Neural Networks (DNNs), have become indispensable for accurate and robust support of large-scale cloud based AI services. However, serving giant DNNs is prohibitively expensive from an energy consumption viewpoint easily exceeding that of training, due to the enormous scale of GPU clusters needed to hold giant DNN model partitions and replicas. Existing approaches can either optimize energy efficiency or inference accuracy but not both. To overcome this status quo, we propose HybridServe, a novel hybrid DNN model serving system that leverages multiple sized versions (small to giant) of the model to be served in tandem. Through a confidence based hybrid model serving dataflow, HybridServe prefers to serve inference requests with energy-efficient smaller models so long as accuracy is not compromised, thereby reducing the number of replicas needed for giant DNNs. HybridServe also features a dataflow planner for efficient partitioning and replication of candidate models to maximize serving system throughput. Experimental results using a prototype implementation of HybridServe show that it reduces energy footprint by up to 19.8x compared to the state-of-the-art DNN model serving systems while matching the accuracy of serving solely with giant DNNs.

cs.LG↗
arXiv2025

MoE-Infinity: Efficient MoE Inference on Personal Machines with Sparsity-Aware Expert Cache

This paper presents MoE-Infinity, an efficient MoE inference system designed for personal machines with limited GPU memory capacity. The key idea for MoE-Infinity is that on personal machines, which are often single-user environments, MoE-based LLMs typically operate with a batch size of one. In this setting, MoE models exhibit a high degree of activation sparsity, meaning a small number of experts are frequently reused in generating tokens during the decode phase. Leveraging this idea, we design a sparsity-aware expert cache, which can trace the sparse activation of experts during inference and carefully select the trace that represents the sparsity pattern. By analyzing these selected traces, MoE-Infinity guides the replacement and prefetching of the expert cache, providing 3.1-16.7x per-token latency improvements over numerous state-of-the-art systems, including vLLM, Ollama, DeepSpeed and BrainStorm across various MoE models (DeepSeek and Mixtral) when handling different LLM tasks. MoE-Infinity's source code is publicly available at https://github.com/EfficientMoE/MoE-Infinity

cs.LG↗
arXiv2025

MoE-Gen: High-Throughput MoE Inference on a Single GPU with Module-Based Batching

This paper presents MoE-Gen, a high-throughput MoE inference system optimized for single-GPU execution. Existing inference systems rely on model-based or continuous batching strategies, originally designed for interactive inference, which result in excessively small batches for MoE's key modules-attention and expert modules-leading to poor throughput. To address this, we introduce module-based batching, which accumulates tokens in host memory and dynamically launches large batches on GPUs to maximize utilization. Additionally, we optimize the choice of batch sizes for each module in an MoE to fully overlap GPU computation and communication, maximizing throughput. Evaluation demonstrates that MoE-Gen achieves 8-31x higher throughput compared to state-of-the-art systems employing model-based batching (FlexGen, MoE-Lightning, DeepSpeed), and offers even greater throughput improvements over continuous batching systems (e.g., vLLM and Ollama) on popular MoE models (DeepSeek and Mixtral) across offline inference tasks. MoE-Gen's source code is publicly available at https://github.com/EfficientMoE/MoE-Gen

cs.DC↗
arXiv2024

Iterative structural coarse-graining for contagion dynamics in complex networks

Contagion dynamics in complex networks drive critical phenomena such as epidemic spread and information diffusion,but their analysis remains computationally prohibitive in large-scale, high-complexity systems. Here, we introduce the Iterative Structural Coarse-Graining (ISCG) framework, a scalable methodology that reduces network complexity while preserving key contagion dynamics with high fidelity. Importantly, we derive theoretical conditions ensuring the precise preservation of both macroscopic outbreak sizes and microscopic node-level infection probabilities during network reduction. Under these conditions, extensive experiments on diverse empirical networks demonstrate that ISCG achieves significant complexity reduction without sacrificing prediction accuracy. Beyond simplification, ISCG reveals multiscale structural patterns that govern contagion processes, enabling practical solutions to longstanding challenges in contagion dynamics. Specifically, ISCG outperforms traditional adaptive centrality-based approaches in identifying influential spreaders, immunizing critical edges, and optimizing sentinel placement for early outbreak detection, offering superior accuracy and computational efficiency. By bridging computational efficiency with dynamical fidelity, ISCG provides a transformative framework for analyzing large-scale contagion processes, with broad applications for epidemiology, information dissemination, and network resilience.

physics.soc-ph↗
arXiv2024

MoE-CAP: Benchmarking Cost, Accuracy and Performance of Sparse Mixture-of-Experts Systems

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently, but it depends on heterogeneous compute and memory resources. These factors jointly affect system Cost, Accuracy, and Performance (CAP), making trade-offs inevitable. Existing benchmarks often fail to capture these trade-offs accurately, complicating practical deployment decisions. To address this, we introduce MoE-CAP, a benchmark specifically designed for MoE systems. Our analysis reveals that achieving an optimal balance across CAP is difficult with current hardware; MoE systems typically optimize two of the three dimensions at the expense of the third-a dynamic we term the MoE-CAP trade-off. To visualize this, we propose the CAP Radar Diagram. We further introduce sparsity-aware performance metrics-Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU)-to enable accurate performance benchmarking of MoE systems across diverse hardware platforms and deployment scenarios.

cs.LG↗
arXiv2024

Social contagion with emotional group interactions

Individual decisions and behaviors are shaped not only by direct interactions with others but also by the collective emotional dynamics within groups. In this work, we introduce the signed simplicial contagion model, integrating both pairwise and emotional group interactions to investigate contagion dynamics in signed networks. Through mean field analysis and numerical simulations, we show that emotional group interactions can induce discontinuous phase transitions, bistable behavior, and hysteresis loops. However, as the proportion of negative edges q increases, the influence of group interactions weakens under a given transmission strength, driving a shift from discontinuous to continuous phase transitions. Our findings reveal that pairwise and group interactions respond differently to changes in q: group interactions display nonlinear sensitivity, while pairwise interactions exhibit a more gradual, linear response. This divergence shifts the dominant mechanisms of contagion, depending on the levels of trust and distrust in the network, providing deeper insights into how emotional relational shape the spread of contagion in social systems.

physics.soc-ph↗
arXiv2024

ServerlessLLM: Low-Latency Serverless Inference for Large Language Models

This paper presents ServerlessLLM, a distributed system designed to support low-latency serverless inference for Large Language Models (LLMs). By harnessing the substantial near-GPU storage and memory capacities of inference servers, ServerlessLLM achieves effective local checkpoint storage, minimizing the need for remote checkpoint downloads and ensuring efficient checkpoint loading. The design of ServerlessLLM features three core contributions: (i) \emph{fast multi-tier checkpoint loading}, featuring a new loading-optimized checkpoint format and a multi-tier loading system, fully utilizing the bandwidth of complex storage hierarchies on GPU servers; (ii) \emph{efficient live migration of LLM inference}, which enables newly initiated inferences to capitalize on local checkpoint storage while ensuring minimal user interruption; and (iii) \emph{startup-time-optimized model scheduling}, which assesses the locality statuses of checkpoints on each server and schedules the model onto servers that minimize the time to start the inference. Comprehensive evaluations, including microbenchmarks and real-world scenarios, demonstrate that ServerlessLLM dramatically outperforms state-of-the-art serverless systems, reducing latency by 10 - 200X across various LLM inference workloads.

cs.LG↗
arXiv2024

Nucleation phenomena and extreme vulnerability of spatial k-core systems

K-core percolation is a fundamental dynamical process in complex networks with applications that span numerous real-world systems. Earlier studies focus primarily on random networks without spatial constraints and reveal intriguing mixed-order transitions. However, real-world systems, ranging from transportation and communication networks to complex brain networks, are not random but are spatially embedded. Here, we study k-core percolation on two-dimensional spatially embedded networks and show that, in contrast to regular percolation, the length of connections can control the transition type, leading to four different types of phase transitions associated with novel phenomena and a rich phase diagram. A key finding is the existence of a metastable phase in which microscopic localized damage, independent of system size, can cause a macroscopic phase transition, a result which cannot be achieved in traditional percolation. In this case, local failures can spontaneously propagate the damage radially until the system entirely collapses, a phenomenon analogous to the nucleation process. These findings suggest novel features and extreme vulnerabilities of spatially embedded k-core network systems, and highlight the necessity to take into account the characteristic length of links when designing robust spatial networks. Furthermore, our insight about the microscopic processes and their origin during the mixed order and first order abrupt transitions in k-core networks could shed light on the mechanisms of many systems where such transitions occur.

physics.soc-ph↗
arXiv2023

Community detection based on structural balance in signed networks

In signed networks, some existing community detection methods treat negative connections as intercommunity links and positive ones as intracommunity links. However, it is important to recognize that negative links on real-world networks also play a key role in maintaining community stability. In this work, our aim is to identify communities that are not only densely connected but also harmonious or balanced in terms of the nature of their relationships. Such communities are more likely to be stable over time and less prone to conflicts. Consequently, we propose a motif-based method to identify communities by quantifying the importance of links in the local structural balance. The results in synthetic and real-world networks show that the proposed method has a higher performance in identifying the community. In addition, it demonstrates strong robustness, i.e., remains insensitive to the balance of the network, and accurately classifies communities in real-world networks.

physics.soc-ph↗
arXiv2023

Possible origin for the similar phase transitions in k-core and interdependent networks

The models of $k$-core percolation and interdependent networks (IN) have been extensively studied in their respective fields. A recent study has revealed that they share several common critical exponents. However, several newly discovered exponents in IN have not been explored in $k$-core percolation, and the origin of the similarity still remains unclear. Here, we investigate k-core percolation in random networks. We find that for k-core percolation,the fractality of the giant component fluctuations is manifested by a fractal fluctuation dimension, $\widetilde d_f = 3/4$, within a correlation \emph{size} $N'$ that scales as $N' \propto (p-p_c)^{-\widetildeν}$, with $\widetildeν= 2$, same as found in IN. Indeed, here, $\widetildeν\equiv d\cdot ν'$ and $\widetilde{d}_f \equiv d'_f/d$, where $ν'$ and $d'_f$ are respectively the same as the correlation \emph{length} exponent and the fractal fluctuation dimension observed in $d$-dimensional IN spatial networks. These two new exponents found here for $k$-core percolation demonstrate the same scaling behaviors as found for IN with the same critical exponents, reinforcing the similarity between the two models. Furthermore, we suggest that these two models are similar since both have two types of interactions: short-range (SR) connectivity and long-range (LR) influences. In IN the LR are the influences of dependency links while in k-core we find here that for $k=1$ and $k=2$ the influences are short range while for $k\geq3$ the influence is long range. In addition, analytical arguments for a universal hyper-scaling relation for the fractal fluctuation dimension of the $k$-core giant component and for IN as well as for any mixed-order transition are established.Our analysis enhances the comprehension of k-core percolation and supports the generalization of the concept of fractal fluctuations in mixed-order phase transitions.

physics.soc-ph↗