arXiv subjects

Yifan Xiong

Publications and source records attributed to Yifan Xiong.

14 recordsLinked to original sources

arXiv2026

ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

In prefill-decode (PD) disaggregated LLM serving, each request is assigned to a decode worker after prefill. Existing decode routers balance only load; for mixture-of-experts (MoE) models this is incomplete: equally loaded workers can differ in latency, since each decode step loads the weights of every distinct expert its batch activates. We present ELDR, an expert-locality-aware decode router for PD-disaggregated MoE serving. From a request's prefill expert activations, ELDR builds an expert signature predicting the experts it will activate during generation. Offline, balanced K-means partitions signature space across decode workers; online, locality-band routing sends each request to the least-loaded worker among those best matching its signature. A signature cache, co-indexed with the KV cache at KV-block granularity, keeps signatures exact under prefix caching. Implemented in vLLM and evaluated on deployments of up to 40 GPUs, ELDR reduces median TPOT by 5.9-13.9% over the strongest of four load-balancing baselines across three MoE models and two workloads, with model outputs unchanged.

cs.DC↗

arXiv2026

TSGuard: Automated User-Centric Incident Diagnosis for AI Workloads in the Cloud

AI workloads incur frequent failures and incidents from the underlying infrastructure. The current incident management workflow follows a provider-centric paradigm, where users report incidents to the infrastructure provider who then conducts troubleshooting. Due to the large number of incidents and the manual nature of the troubleshooting process, the provider often takes several days to resolve an incident, resulting in operational delays and productivity loss. To address these challenges, we present TSGuard, a user-centric multi-agent system that delivers immediate incident diagnosis to users who deploy the workloads. The core innovation of TSGuard is twofold: (1) constructing domain-specific knowledge bases by mining historical on-call experiences in the offline phase, and (2) mimicking human expert diagnosis via structured reasoning and iterative trial-and-error in the online phase. Evaluation using production incident records from Microsoft Azure demonstrates that TSGuard significantly outperforms state-of-the-art baselines, improving diagnostic accuracy by 19.8%. Furthermore, TSGuard reduces the average verification time by 63.4% compared to the sequential execution baseline.

cs.SE↗

arXiv2026

Benchmarking AI scientists for omics data driven biological discovery

Recent advances in large language models have enabled the emergence of AI scientists that aim to autonomously analyze biological data and assist scientific discovery. Despite rapid progress, it remains unclear to what extent these systems can extract meaningful biological insights from real experimental data. Existing benchmarks either evaluate reasoning in the absence of data or focus on predefined analytical outputs, failing to reflect realistic, data-driven biological research. Here, we introduce BAISBench (Biological AI Scientist Benchmark), a benchmark for evaluating AI scientists on real single-cell transcriptomic datasets. BAISBench comprises two tasks: cell type annotation across 15 expert-labeled datasets, and scientific discovery through 193 multiple-choice questions derived from biological conclusions reported in 41 published single-cell studies. We evaluated several representative AI scientists using BAISBench and, to provide a human performance baseline, invited six graduate-level bioinformaticians to collectively complete the same tasks. The results show that while current AI scientists fall short of fully autonomous biological discovery, they already demonstrate substantial potential in supporting data-driven biological research. These results position BAISBench as a practical benchmark for characterizing the current capabilities and limitations of AI scientists in biological research. We expect BAISBench to serve as a practical evaluation framework for guiding the development of more capable AI scientists and for helping biologists identify AI systems that can effectively support real-world research workflows. The BAISBench can be found at: https://github.com/EperLuo/BAISBench, https://huggingface.co/datasets/EperLuo/BaisBench.

cs.AI↗

arXiv2025

Sigma-MoE-Tiny Technical Report

Mixture-of-Experts (MoE) has emerged as a promising paradigm for foundation models due to its efficient and powerful scalability. In this work, we present Sigma-MoE-Tiny, an MoE language model that achieves the highest sparsity compared to existing open-source models. Sigma-MoE-Tiny employs fine-grained expert segmentation with up to 96 experts per layer, while activating only one expert for each token, resulting in 20B total parameters with just 0.5B activated. The major challenge introduced by such extreme sparsity lies in expert load balancing. We find that the widely-used load balancing loss tends to become ineffective in the lower layers under this setting. To address this issue, we propose a progressive sparsification schedule aiming to balance expert utilization and training stability. Sigma-MoE-Tiny is pre-trained on a diverse and high-quality corpus, followed by post-training to further unlock its capabilities. The entire training process remains remarkably stable, with no occurrence of irrecoverable loss spikes. Comprehensive evaluations reveal that, despite activating only 0.5B parameters, Sigma-MoE-Tiny achieves top-tier performance among counterparts of comparable or significantly larger scale. In addition, we provide an in-depth discussion of load balancing in highly sparse MoE models, offering insights for advancing sparsity in future MoE architectures. Project page: https://qghuxmu.github.io/Sigma-MoE-Tiny Code: https://github.com/microsoft/ltp-megatron-lm

cs.CL↗

arXiv2025

SIGMA: An AI-Empowered Training Stack on Early-Life Hardware

An increasing variety of AI accelerators is being considered for large-scale training. However, enabling large-scale training on early-life AI accelerators faces three core challenges: frequent system disruptions and undefined failure modes that undermine reliability; numerical errors and training instabilities that threaten correctness and convergence; and the complexity of parallelism optimization combined with unpredictable local noise that degrades efficiency. To address these challenges, SIGMA is an open-source training stack designed to improve the reliability, stability, and efficiency of large-scale distributed training on early-life AI hardware. The core of this initiative is the LUCIA TRAINING PLATFORM (LTP), the system optimized for clusters with early-life AI accelerators. Since its launch in March 2025, LTP has significantly enhanced training reliability and operational productivity. Over the past five months, it has achieved an impressive 94.45% effective cluster accelerator utilization, while also substantially reducing node recycling and job-recovery times. Building on the foundation of LTP, the LUCIA TRAINING FRAMEWORK (LTF) successfully trained SIGMA-MOE, a 200B MoE model, using 2,048 AI accelerators. This effort delivered remarkable stability and efficiency outcomes, achieving 21.08% MFU, state-of-the-art downstream accuracy, and encountering only one stability incident over a 75-day period. Together, these advances establish SIGMA, which not only tackles the critical challenges of large-scale training but also establishes a new benchmark for AI infrastructure and platform innovation, offering a robust, cost-effective alternative to prevailing established accelerator stacks and significantly advancing AI capabilities and scalability. The source code of SIGMA is available at https://github.com/microsoft/LuciaTrainingPlatform.

cs.DC↗

arXiv2025

Aerial Grasping via Maximizing Delta-Arm Workspace Utilization

The workspace limits the operational capabilities and range of motion for the systems with robotic arms. Maximizing workspace utilization has the potential to provide more optimal solutions for aerial manipulation tasks, increasing the system's flexibility and operational efficiency. In this paper, we introduce a novel planning framework for aerial grasping that maximizes workspace utilization. We formulate an optimization problem to optimize the aerial manipulator's trajectory, incorporating task constraints to achieve efficient manipulation. To address the challenge of incorporating the delta arm's non-convex workspace into optimization constraints, we leverage a Multilayer Perceptron (MLP) to map position points to feasibility probabilities.Furthermore, we employ Reversible Residual Networks (RevNet) to approximate the complex forward kinematics of the delta arm, utilizing efficient model gradients to eliminate workspace constraints. We validate our methods in simulations and real-world experiments to demonstrate their effectiveness.

cs.RO↗

arXiv2025

Argos: Agentic Time-Series Anomaly Detection with Autonomous Rule Generation via Large Language Models

Observability in cloud infrastructure is critical for service providers, driving the widespread adoption of anomaly detection systems for monitoring metrics. However, existing systems often struggle to simultaneously achieve explainability, reproducibility, and autonomy, which are three indispensable properties for production use. We introduce Argos, an agentic system for detecting time-series anomalies in cloud infrastructure by leveraging large language models (LLMs). Argos proposes to use explainable and reproducible anomaly rules as intermediate representation and employs LLMs to autonomously generate such rules. The system will efficiently train error-free and accuracy-guaranteed anomaly rules through multiple collaborative agents and deploy the trained rules for low-cost online anomaly detection. Through evaluation results, we demonstrate that Argos outperforms state-of-the-art methods, increasing $F_1$ scores by up to $9.5\%$ and $28.3\%$ on public anomaly detection datasets and an internal dataset collected from Microsoft, respectively.

cs.LG↗

arXiv2024

Attentive-based Multi-level Feature Fusion for Voice Disorder Diagnosis

Voice disorders negatively impact the quality of daily life in various ways. However, accurately recognizing the category of pathological features from raw audio remains a considerable challenge due to the limited dataset. A promising method to handle this issue is extracting multi-level pathological information from speech in a comprehensive manner by fusing features in the latent space. In this paper, a novel framework is designed to explore the way of high-quality feature fusion for effective and generalized detection performance. Specifically, the proposed model follows a two-stage training paradigm: (1) ECAPA-TDNN and Wav2vec 2.0 which have shown remarkable effectiveness in various domains are employed to learn the universal pathological information from raw audio; (2) An attentive fusion module is dedicatedly designed to establish the interaction between pathological features projected by EcapTdnn and Wav2vec 2.0 respectively and guide the multi-layer fusion, the entire model is jointly fine-tuned from pre-trained features by the automatic voice pathology detection task. Finally, comprehensive experiments on the FEMH and SVD datasets demonstrate that the proposed framework outperforms the competitive baselines, and achieves the accuracy of 90.51% and 87.68%.

cs.SD↗

arXiv2024

Demystifying and Assessing Code Understandability in Java Decompilation

Decompilation, the process of converting machine-level code into readable source code, plays a critical role in reverse engineering. Given that the main purpose of decompilation is to facilitate code comprehension in scenarios where the source code is unavailable, the understandability of decompiled code is of great importance. In this paper, we propose the first empirical study on the understandability of Java decompiled code and obtained the following findings: (1) Understandability of Java decompilation is considered as important as its correctness, and decompilation understandability issues are even more commonly encountered than decompilation failures. (2) A notable percentage of code snippets decompiled by Java decompilers exhibit significantly lower or higher levels of understandability in comparison to their original source code. (3) Unfortunately, Cognitive Complexity demonstrates relatively acceptable precision while low recall in recognizing these code snippets exhibiting diverse understandability during decompilation. (4) Even worse, perplexity demonstrates lower levels of precision and recall in recognizing such code snippets. Inspired by the four findings, we further proposed six code patterns and the first metric for the assessment of decompiled code understandability. This metric was extended from Cognitive Complexity, with six more rules harvested from an exhaustive manual analysis into 1287 pairs of source code snippets and corresponding decompiled code. This metric was also validated using the original and updated dataset, yielding an impressive macro F1-score of 0.88 on the original dataset, and 0.86 on the test set.

cs.SE↗

arXiv2024

Four-fold truncated double-nested anti-resonant hollow-core fibers with ultralow loss and ultrahigh mode purity

Hollow-core fibers are inherently multimode, making it crucial to filter out higher-order modes within the shortest possible fiber length for applications such as high speed coherent communications and fiber optic gyroscopes. However, current HCF designs face the challenges of simultaneously achieving ultralow fundamental mode loss and ultrahigh HOM suppression. In this study, we present a novel four fold truncated double nested anti resonant hollow core fiber structure that addresses this challenge. Our 4T-DNANF enables greater control over phase-matching between core modes and air modes in the cladding, allowing for minimized FM loss and substantially increased HOM loss. Experimentally, we fabricated several HCFs: one with an FM loss of 0.1 dB/km and an HOM loss of 430 dB/km, and another with an FM loss of 0.13 dB/km with a HOM loss of 6500 dB/km, resulting in a higher-order mode extinction ratio of 50,000.

physics.optics↗

arXiv2024

SuperBench: Improving Cloud AI Infrastructure Reliability with Proactive Validation

Reliability in cloud AI infrastructure is crucial for cloud service providers, prompting the widespread use of hardware redundancies. However, these redundancies can inadvertently lead to hidden degradation, so called "gray failure", for AI workloads, significantly affecting end-to-end performance and concealing performance issues, which complicates root cause analysis for failures and regressions. We introduce SuperBench, a proactive validation system for AI infrastructure that mitigates hidden degradation caused by hardware redundancies and enhances overall reliability. SuperBench features a comprehensive benchmark suite, capable of evaluating individual hardware components and representing most real AI workloads. It comprises a Validator which learns benchmark criteria to clearly pinpoint defective components. Additionally, SuperBench incorporates a Selector to balance validation time and issue-related penalties, enabling optimal timing for validation execution with a tailored subset of benchmarks. Through testbed evaluation and simulation, we demonstrate that SuperBench can increase the mean time between incidents by up to 22.61x. SuperBench has been successfully deployed in Azure production, validating hundreds of thousands of GPUs over the last two years.

cs.DC↗

arXiv2023

FP8-LM: Training FP8 Large Language Models

In this paper, we explore FP8 low-bit data formats for efficient training of large language models (LLMs). Our key insight is that most variables, such as gradients and optimizer states, in LLM training can employ low-precision data formats without compromising model accuracy and requiring no changes to hyper-parameters. Specifically, we propose a new FP8 automatic mixed-precision framework for training LLMs. This framework offers three levels of FP8 utilization to streamline mixed-precision and distributed parallel training for LLMs. It gradually incorporates 8-bit gradients, optimizer states, and distributed learning in an incremental manner. Experiment results show that, during the training of GPT-175B model on H100 GPU platform, our FP8 mixed-precision training framework not only achieved a remarkable 39% reduction in real memory usage but also ran 75% faster than the widely adopted BF16 framework (i.e., Megatron-LM), surpassing the speed of Nvidia Transformer Engine by 37%. This largely reduces the training costs for large foundation models. Furthermore, our FP8 mixed-precision training methodology is generic. It can be seamlessly applied to other tasks such as LLM instruction tuning and reinforcement learning with human feedback, offering savings in fine-tuning expenses. Our FP8 low-precision training framework is open-sourced at {https://github.com/Azure/MS-AMP}{aka.ms/MS.AMP}.

cs.LG↗

arXiv2023

Tutel: Adaptive Mixture-of-Experts at Scale

Sparsely-gated mixture-of-experts (MoE) has been widely adopted to scale deep learning models to trillion-plus parameters with fixed computational cost. The algorithmic performance of MoE relies on its token routing mechanism that forwards each input token to the right sub-models or experts. While token routing dynamically determines the amount of expert workload at runtime, existing systems suffer inefficient computation due to their static execution, namely static parallelism and pipelining, which does not adapt to the dynamic workload. We present Flex, a highly scalable stack design and implementation for MoE with dynamically adaptive parallelism and pipelining. Flex designs an identical layout for distributing MoE model parameters and input data, which can be leveraged by all possible parallelism or pipelining methods without any mathematical inequivalence or tensor migration overhead. This enables adaptive parallelism/pipelining optimization at zero cost during runtime. Based on this key design, Flex also implements various MoE acceleration techniques. Aggregating all techniques, Flex finally delivers huge speedup at any scale -- 4.96x and 5.75x speedup of a single MoE layer over 16 and 2,048 A100 GPUs, respectively, over the previous state-of-the-art. Our evaluation shows that Flex efficiently and effectively runs a real-world MoE-based model named SwinV2-MoE, built upon Swin Transformer V2, a state-of-the-art computer vision architecture. On efficiency, Flex accelerates SwinV2-MoE, achieving up to 1.55x and 2.11x speedup in training and inference over Fairseq, respectively. On effectiveness, the SwinV2-MoE model achieves superior accuracy in both pre-training and down-stream computer vision tasks such as COCO object detection than the counterpart dense model, indicating the readiness of Flex for end-to-end real-world model training and inference.

cs.DC↗

arXiv2022

GC3: An Optimizing Compiler for GPU Collective Communication

Machine learning models made up of millions or billions of parameters are trained and served on large multi-GPU systems. As models grow in size and execute on more GPUs, the collective communications used in these applications become a bottleneck. Custom collective algorithms optimized for both particular network topologies and application specific communication patterns can alleviate this bottleneck and help these applications scale. However, correctly and efficiently implementing custom algorithms is challenging. This paper introduces GC3, a system for programmable GPU communication. GC3 provides a domain specific language for writing collective communication algorithms and an optimizing compiler for lowering them to an executable form, which can be executed efficiently and flexibly in an interpreter based runtime. We used GC3 to write novel collective algorithms for AllReduce and AllToAll that are up to $1.9\times$ and $1.3\times$ faster than hand-optimized implementations, respectively.

cs.DC↗