arXiv subjects

Yi Xiong

Publications and source records attributed to Yi Xiong.

At least 19 recordsLinked to original sources

arXiv2026

Source-Prior-Driven Selective Adaptation for Efficient Diffusion Model Finetuning

Fine-tuning large diffusion models for new domains or styles involves a trade-off: improving target-specific generation often degrades the pretrained model's broad generative capability. Existing full and parameter-efficient fine-tuning methods typically handle this trade-off only implicitly. In this work, we propose a novel source-prior-driven selective adaptation method to efficiently fine-tune diffusion models, achieving a favorable trade-off. Our method relies on two key observations: (1) the loss of general generative capability is highly inconsistent across pretrained parameters, and (2) parameters that have a relatively small impact on the model's general generative capability remain structurally inconsistent across layers and parameter types. Motivated by these observations, we first learn a static mask to explicitly identify parameters better suited for downstream adaptation, and then construct structured update strategies for the selected subset. Experiments show that our method achieves a better adaptation-retention trade-off than existing strong baselines.

cs.AI↗

arXiv2026

eLLM: Elastic Memory Management Framework for Efficient LLM Serving

Large Language Models are increasingly being deployed in datacenters. Serving these models requires careful memory management, as their memory usage includes static weights, dynamic activations, and key-value caches. While static weights are constant and predictable, dynamic components such as activations and KV caches change frequently during runtime, presenting significant challenges for efficient memory management. Modern LLM serving systems typically handle runtime memory and KV caches at distinct abstraction levels: runtime memory management relies on static tensor abstractions, whereas KV caches utilize a page table-based virtualization layer built on top of the tensor abstraction. This virtualization dynamically manages KV caches to mitigate memory fragmentation. However, this dual-level approach fundamentally isolates runtime memory and KV cache management, resulting in suboptimal memory utilization under dynamic workloads, which can lead to a nearly 20% drop in throughput. To address these limitations, we propose eLLM, an elastic memory management framework inspired by the classical memory ballooning mechanism in operating systems. The core components of eLLM include: (1) Virtual Tensor Abstraction, which decouples the virtual address space of tensors from the physical GPU memory, creating a unified and flexible memory pool; (2) an Elastic Memory Mechanism that dynamically adjusts memory allocation through runtime memory inflation and deflation, leveraging CPU memory as an extensible buffer; and (3) a Lightweight Scheduling Strategy employing SLO-aware policies to optimize memory utilization and effectively balance performance trade-offs under stringent SLO constraints. Comprehensive evaluations demonstrate that eLLM significantly outperforms state-of-the-art systems, 2.32x higher decoding throughput, and supporting 3x larger batch sizes for 128K-token inputs.

cs.DC↗

arXiv2026

Scaffold-Conditioned Preference Triplets for Controllable Molecular Optimization with Large Language Models

Molecular property optimization is central to drug discovery, yet many deep learning methods rely on black-box scoring and offer limited control over scaffold preservation, often producing unstable or biologically implausible edits. While large language models (LLMs) are promising molecular generators, optimization remains constrained by the lack of chemistry-grounded preference supervision and principled data curation. We introduce \textbf{Scaffold-Conditioned Preference Triplets (SCPT)}, a pipeline that constructs similarity-constrained triplets $\langle\text{scaffold}, \text{better}, \text{worse}\rangle$ via scaffold alignment and chemistry-driven filters for validity, synthesizability, and meaningful property gains. Using these preferences, we align a pretrained molecular LLM as a conditional editor, enabling property-improving edits that retain the scaffold. Across single- and multi-objective benchmarks, SCPT improves optimization success and property gains while maintaining higher scaffold similarity than competitive baselines. Compared with representative non-LLM molecular optimization methods, SCPT-trained LLMs are better suited to scaffold-constrained and multi-objective optimization. In addition, models trained on single-property and two-property supervision generalize effectively to three-property tasks, indicating promising extrapolative generalization under limited higher-order supervision. SCPT also provides controllable data-construction knobs that yield a predictable similarity-gain frontier, enabling systematic adaptation to diverse optimization regimes.

cs.LG↗

arXiv2026

Age-Specific Logistic Regression with Complex Event Time Data

In attempt to advance the current practice for assessing and predicting the primary ovarian insufficiency (POI) risk in female childhood cancer survivors, we propose two estimating function based approaches for age-specific logistic regression. Both approaches adapt the inverse probability of censoring weighting (IPCW) strategy and yield consistent estimators with asymptotic normality. The first approach modifies the IPCW weights used by Im et al. (2023) to account for doubly censoring. The second approach extends the outcome weighted IPCW approach to use the information of the subjects censored before the analysis time. We consider variance estimation for the estimators and explore by simulation the two approaches implemented in the situations where the conditional right-censoring time distribution required in the IPCW weighs is unknown and approximated using the survival random forest approaches, stratified empirical distribution functions, or the estimator under the Cox proportional hazards model. The numerical studies indicate that the second approach is more efficient when right-censoring is relatively heavy, whereas the first approach is preferable when the right-censoring is light. We also observe that the performance of the two approaches heavily relies on the estimation of censoring distribution in our simulation settings. The POI data from a childhood cancer survivor study are employed throughout the paper for motivation and illustration. Our data analysis provides new insight into understanding the POI risk among cancer survivors.

stat.ME↗

arXiv2026

Consistency-Driven Dual LSTM Models for Kinematic Control of a Wearable Soft Robotic Arm

In this paper, we introduce a consistency-driven dual LSTM framework for accurately learning both the forward and inverse kinematics of a pneumatically actuated soft robotic arm integrated into a wearable device. This approach effectively captures the nonlinear and hysteretic behaviors of soft pneumatic actuators while addressing the one-to-many mapping challenge between actuation inputs and end-effector positions. By incorporating a cycle consistency loss, we enhance physical realism and improve the stability of inverse predictions. Extensive experiments-including trajectory tracking, ablation studies, and wearable demonstrations-confirm the effectiveness of our method. Results indicate that the inclusion of the consistency loss significantly boosts prediction accuracy and promotes physical consistency over conventional approaches. Moreover, the wearable soft robotic arm demonstrates strong human-robot collaboration capabilities and adaptability in everyday tasks such as object handover, obstacle-aware pick-and-place, and drawer operation. This work underscores the promising potential of learning-based kinematic models for human-centric, wearable robotic systems.

cs.RO↗

arXiv2026

Artificial Intelligence for Detecting Fetal Orofacial Clefts and Advancing Medical Education

Orofacial clefts are among the most common congenital craniofacial abnormalities, yet accurate prenatal detection remains challenging due to the scarcity of experienced specialists and the relative rarity of the condition. Early and reliable diagnosis is essential to enable timely clinical intervention and reduce associated morbidity. Here we show that an artificial intelligence system, trained on over 45,139 ultrasound images from 9,215 fetuses across 22 hospitals, can diagnose fetal orofacial clefts with sensitivity and specificity exceeding 93% and 95% respectively, matching the performance of senior radiologists and substantially outperforming junior radiologists. When used as a medical copilot, the system raises junior radiologists' sensitivity by more than 6%. Beyond direct diagnostic assistance, the system also accelerates the development of clinical expertise. A pilot study involving 24 radiologists and trainees demonstrated that the model can improve the expertise development for rare conditions. This dual-purpose approach offers a scalable solution for improving both diagnostic accuracy and specialist training in settings where experienced radiologists are scarce.

cs.CV↗

arXiv2025

Is Thompson Sampling Susceptible to Algorithmic Collusion?

When two players are engaged in a repeated game with unknown payoff matrices, they may use single-agent multi-armed bandit algorithms to choose the actions independent of each other. We show that when the players use Thompson sampling, the game dynamics converges to the Nash equilibrium under a mild assumption on the payoff matrices. Therefore, algorithmic collusion doesn't arise in this case despite the fact that the players do not intentionally deploy competitive strategies. To prove the convergence result, we find that the framework developed in stochastic approximation doesn't apply, because of the sporadic and infrequent updates of the inferior actions and the lack of Lipschitz continuity. We develop a novel sample-path-wise approach to show the convergence. However, when the payoff matrices do not satisfy the assumption, the game may converge to collusive outcomes.

cs.GT↗

arXiv2025

FlashRecovery: Fast and Low-Cost Recovery from Failures for Large-Scale Training of LLMs

Large language models (LLMs) have made a profound impact across various fields due to their advanced capabilities. However, training these models at unprecedented scales requires extensive AI accelerator clusters and sophisticated parallelism strategies, which pose significant challenges in maintaining system reliability over prolonged training periods. A major concern is the substantial loss of training time caused by inevitable hardware and software failures. To address these challenges, we present FlashRecovery, a fast and low-cost failure recovery system comprising three core modules: (1) Active and real-time failure detection. This module performs continuous training state monitoring, enabling immediate identification of hardware and software failures within seconds, thus ensuring rapid incident response; (2) Scale-independent task restart. By employing different recovery strategies for normal and faulty nodes, combined with an optimized communication group reconstruction protocol, our approach ensures that the recovery time remains nearly constant, regardless of cluster scale; (3) Checkpoint-free recovery within one step. Our novel recovery mechanism enables single-step restoration, completely eliminating dependence on traditional checkpointing methods and their associated overhead. Collectively, these innovations enable FlashRecovery to achieve optimal Recovery Time Objective (RTO) and Recovery Point Objective (RPO), substantially improving the reliability and efficiency of long-duration LLM training. Experimental results demonstrate that FlashRecovery system can achieve training restoration on training cluster with 4, 800 devices in 150 seconds. We also verify that the time required for failure recovery is nearly consistent for different scales of training tasks.

cs.DC↗

arXiv2025

Investigating the shadows of new regular black holes with a Minkowski core: Effects of spherical accretion and core type differences

We investigated the shadows and optical appearances of a new type of regular black holes (BHs) with a Minkowski core under various spherical accretion scenarios. These BHs are constructed by modifying the Newtonian potential based on the minimum observable length in the Generalized Uncertainty Principle (GUP). They correspond one-to-one with traditional regular BHs featuring a de-Sitter (dS) core (such as Bardeen/Hayward BHs), characterized by a quantum gravity effect parameter ($α_0$) and spacetime deformation factor ($n$). We found that the characteristic parameters give rise to some novel observable features. For these new BHs, both the shadow and photon sphere radii decrease with the increase in $α_0$, while the observed specific intensity increases. Conversely, as n increases, the shadow and photon sphere radii increase, while the observed specific intensity decreases. Under different spherical accretion scenarios, the shadows and photon sphere radii remain identical; however, the observed specific intensity is greater under static spherical accretion than under infalling spherical accretion. Additionally, we found that these regular BHs with different cores exhibit variations in shadows and optical appearances, particularly under static spherical accretion. Compared with Bardeen BH, the new BHs exhibit a lower observed specific intensity, a dimmer photon ring, and smaller shadow and photon sphere radii. Larger values of $α_0$ lead to more significant differences, and a similar trend was also observed when comparing with Hayward BH. Under infalling spherical accretion, the regular BHs with different cores exhibit only slight differences in observed specific intensity, which become more evident when $α_0$ is relatively large.

gr-qc↗

arXiv2025

Serving Large Language Models on Huawei CloudMatrix384

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910 NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s per NPU even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

cs.DC↗

arXiv2025

WindVE: Collaborative CPU-NPU Vector Embedding

Retrieval-Augmented Generation is a technology that enhances large language models by integrating information retrieval. In the industry, inference services based on LLMs are highly sensitive to cost-performance ratio, prompting the need for improving hardware resource utilization in the inference service. Specifically, vector embedding and retrieval processes take up to 20% of the total latency. Therefore, optimizing the utilization of computational resources in vector embeddings is crucial for enhancing the cost-performance ratio of inference processes, which in turn boosts their product competitiveness.In this paper, we analyze the deployment costs of vector embedding technology in inference services, propose a theoretical formula, and determine through the mathematical expression that increasing the capacity to process concurrent queries is the key to reducing the deployment costs of vector embeddings. Therefore, in this paper, we focus on improving the product's capability to process concurrent queries. To optimize concurrency without sacrificing performance, we have designed a queue manager that adeptly offloads CPU peak queries. This manager utilizes a linear regression model to ascertain the optimal queue depths, a critical parameter that significantly influences the efficacy of the system. We further develop a system named WindVE that uses a CPU-NPU heterogeneous architecture to offload peak concurrent queries, which leverages the performance differences between the two processors to effectively manage traffic surges. Through experiments, we compare WindVE to the state-of-the-art vector embedding framework FlagEmbedding, and achieve a concurrency level up to 22.3% higher than the scheme without offloading.

cs.DC↗

arXiv2025

SLO-Aware Scheduling for Large Language Model Inferences

Large language models (LLMs) have revolutionized applications such as code completion, chatbots, and online classification. To elevate user experiences, service level objectives (SLOs) serve as crucial benchmarks for assessing inference services capabilities. In practice, an inference service processes multiple types of tasks, each with its own distinct SLO. To ensure satisfactory user experiences, each request's distinct SLOs should be considered in scheduling. However, existing designs lack this consideration, leading to insufficient hardware utility and suboptimal performance. This paper analyzes scenarios to process tasks with varying SLOs, and introduces a simulated annealing-based scheduler to decide request priority sequence based on a request's SLO, input lengths, and possible output lengths. As the first specialized scheduler for multi-SLO scenarios, this work improves SLO attainment by up to 5x and reduces average latency by 31.6% on Python-Code-23k-ShareGPT and ShareGPT_Vicuna_unfiltered datasets, compared to current state-of-the-art framework vLLM and a new framework LMDeploy.

cs.DC↗

arXiv2025

UB-Mesh: a Hierarchically Localized nD-FullMesh Datacenter Network Architecture

As the Large-scale Language Models (LLMs) continue to scale, the requisite computational power and bandwidth escalate. To address this, we introduce UB-Mesh, a novel AI datacenter network architecture designed to enhance scalability, performance, cost-efficiency and availability. Unlike traditional datacenters that provide symmetrical node-to-node bandwidth, UB-Mesh employs a hierarchically localized nD-FullMesh network topology. This design fully leverages the data locality of LLM training, prioritizing short-range, direct interconnects to minimize data movement distance and reduce switch usage. Although UB-Mesh's nD-FullMesh topology offers several theoretical advantages, its concrete architecture design, physical implementation and networking system optimization present new challenges. For the actual construction of UB-Mesh, we first design the UB-Mesh-Pod architecture, which is based on a 4D-FullMesh topology. UB-Mesh-Pod is implemented via a suite of hardware components that serve as the foundational building blocks, including specifically-designed NPU, CPU, Low-Radix-Switch (LRS), High-Radix-Switch (HRS), NICs and others. These components are interconnected via a novel Unified Bus (UB) technique, which enables flexible IO bandwidth allocation and hardware resource pooling. For networking system optimization, we propose advanced routing mechanism named All-Path-Routing (APR) to efficiently manage data traffic. These optimizations, combined with topology-aware performance enhancements and robust reliability measures like 64+1 backup design, result in 2.04x higher cost-efficiency, 7.2% higher network availability compared to traditional Clos architecture and 95%+ linearity in various LLM training tasks.

cs.AR↗

arXiv2025

High-Throughput LLM inference on Heterogeneous Clusters

Nowadays, many companies possess various types of AI accelerators, forming heterogeneous clusters. Efficiently leveraging these clusters for high-throughput large language model (LLM) inference services can significantly reduce costs and expedite task processing. However, LLM inference on heterogeneous clusters presents two main challenges. Firstly, different deployment configurations can result in vastly different performance. The number of possible configurations is large, and evaluating the effectiveness of a specific setup is complex. Thus, finding an optimal configuration is not an easy task. Secondly, LLM inference instances within a heterogeneous cluster possess varying processing capacities, leading to different processing speeds for handling inference requests. Evaluating these capacities and designing a request scheduling algorithm that fully maximizes the potential of each instance is challenging. In this paper, we propose a high-throughput inference service system on heterogeneous clusters. First, the deployment configuration is optimized by modeling the resource amount and expected throughput and using the exhaustive search method. Second, a novel mechanism is proposed to schedule requests among instances, which fully considers the different processing capabilities of various instances. Extensive experiments show that the proposed scheduler improves throughput by 122.5% and 33.6% on two heterogeneous clusters, respectively.

cs.DC↗

arXiv2025

Every FLOP Counts: Scaling a 300B Mixture-of-Experts LING LLM without Premium GPUs

In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled Bǎilíng in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI.

cs.LG↗

arXiv2024

Pro-PRIME: A general Temperature-Guided Language model to engineer enhanced Stability and Activity in Proteins

Designing protein mutants of both high stability and activity is a critical yet challenging task in protein engineering. Here, we introduce PRIME, a deep learning model, which can suggest protein mutants of improved stability and activity without any prior experimental mutagenesis data of the specified protein. Leveraging temperature-aware language modeling, PRIME demonstrated superior predictive power compared to current state-of-the-art models on the public mutagenesis dataset over 283 protein assays. Furthermore, we validated PRIME's predictions on five proteins, examining the top 30-45 single-site mutations' impact on various protein properties, including thermal stability, antigen-antibody binding affinity, and the ability to polymerize non-natural nucleic acid or resilience to extreme alkaline conditions. Remarkably, over 30% of the AI-recommended mutants exhibited superior performance compared to their pre-mutation counterparts across all proteins and desired properties. Moreover, we have developed an efficient, and successful method based on PRIME to rapidly obtain multi-site mutants with enhanced activity and stability. Hence, PRIME demonstrates the general applicability in protein engineering.

q-bio.BM↗

arXiv2024

LayerKV: Optimizing Large Language Model Serving with Layer-wise KV Cache Management

The expanding context windows in large language models (LLMs) have greatly enhanced their capabilities in various applications, but they also introduce significant challenges in maintaining low latency, particularly in Time to First Token (TTFT). This paper identifies that the sharp rise in TTFT as context length increases is predominantly driven by queuing delays, which are caused by the growing demands for GPU Key-Value (KV) cache allocation clashing with the limited availability of KV cache blocks. To address this issue, we propose LayerKV, a simple yet effective plug-in method that effectively reduces TTFT without requiring additional hardware or compromising output performance, while seamlessly integrating with existing parallelism strategies and scheduling techniques. Specifically, LayerKV introduces layer-wise KV block allocation, management, and offloading for fine-grained control over system memory, coupled with an SLO-aware scheduler to optimize overall Service Level Objectives (SLOs). Comprehensive evaluations on representative models, ranging from 7B to 70B parameters, across various GPU configurations, demonstrate that LayerKV improves TTFT latency up to 69x and reduces SLO violation rates by 28.7%, significantly enhancing the user experience.

cs.DC↗

arXiv2024

Cost-Effectiveness Analysis for Disease Prevention -- A Case Study on Colorectal Cancer Screening

Cancer Screening has been widely recognized as an effective strategy for preventing the disease. Despite its effectiveness, determining when to start screening is complicated, because starting too early increases the number of screenings over lifetime and thus costs but starting too late may miss the cancer that could have been prevented. Therefore, to make an informed recommendation on the age to start screening, it is necessary to conduct cost-effectiveness analysis to assess the gain in life years relative to the cost of screenings. As more large-scale observational studies become accessible, there is growing interest in evaluating cost-effectiveness based on empirical evidence. In this paper, we propose a unified measure for evaluating cost-effectiveness and a causal analysis for the continuous intervention of screening initiation age, under the multi-state modeling with semi-competing risks. Extensive simulation results show that the proposed estimators perform well in realistic scenarios. We perform a cost-effectiveness analysis of the colorectal cancer screening, utilizing data from the large-scale Women's Health Initiative. Our analysis reveals that initiating screening at age 50 years yields the highest quality-adjusted life years with an acceptable incremental cost-effectiveness ratio compared to no screening, providing real-world evidence in support of screening recommendation for colorectal cancer.

stat.ME↗