arXiv subjects

Xuhui Chen

Publications and source records attributed to Xuhui Chen.

At least 19 recordsLinked to original sources

arXiv2026

SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orchestration, and low-level kernel execution. The resulting system achieves 34.22% Model FLOPs Utilization (MFU) with a 2.93x improvement over the open-source baseline recipe while maintaining training stability. Building on this optimized infrastructure, we further establish a CPT and SFT workflow for complex Operations Research (OR) tasks. We refer to the integrated framework as SLAI T-Rex. Using DeepSeek-V4-Flash, we develop OR-oriented CPT and SFT data pipelines that combine collected domain resources with solver-verified synthetic optimization documents. The resulting dataset contains 10K high-quality SFT samples spanning four task categories and three problem representations. The specialized model achieves the highest average zero-shot Pass@1 score among the evaluated models, reaching 71.81% and outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash model by 3.98 and 11.27 percentage points, respectively. Overall, this work demonstrates a full-stack pathway from efficient trillion-parameter model post-training on Ascend infra to domain-specialized Flash models for solver-grounded mathematical modeling, advancing frontier-model systems for complex reasoning.

cs.CL↗

arXiv2026

T-GINEE: A Tensor-Based Multilayer Graph Representation Learning

Traditional network analysis focuses on single-layer networks, real-world systems often form multilayer networks with multiple relationship types. However, existing methods typically fail to capture complex inter-layer dependencies by treating layers independently or aggregating them. To address this, we propose T-GINEE (Tensor-Based Generalized Multilayer-graph Estimating Equation), a statistical regularization framework combining tensor-based generalized estimating equations with task-specific loss to model cross-network correlations explicitly. Key innovations include: (1) CP tensor decomposition capturing structural dependencies via shared latent factors; (2) a generalized estimating equation framework modeling inter-layer correlations through working covariance matrices; and (3) a flexible link function accommodating characteristics like sparsity. Our theoretical analysis establishes consistency and asymptotic normality under mild conditions. Extensive experiments on synthetic and real-world datasets validate T-GINEE's effectiveness for multilayer network analysis.

cs.LG↗

arXiv2026

Metric--Phase Fields: Decoupling Distance and Sign for Thin-Structure Reconstruction from Unoriented Point Clouds

Neural Signed Distance Functions (SDFs) excel at reconstructing watertight manifolds but fail on thin structures and open boundaries due to strict inside--outside constraints. Conversely, Unsigned Distance Fields (UDFs) accommodate general geometries but suffer from gradient singularities at the zero-level set, hindering optimization and extraction. We introduce Metric--Phase Fields (MPFs), a decoupled implicit representation that separates metric proximity from topological phase. Given an unoriented point cloud, MPFs learn (i) an unsigned metric field $r$ and (ii) a smooth phase field $\theta$, for which we derive a bounded phase indicator $P=\tanh(\beta\theta)$ that provides soft inside--outside cues where they are meaningful. We couple the two fields via a gated-metric formulation with a residual phase injection to obtain a signed implicit function with stable near-surface gradients. The phase coefficient $\beta$ is learnable, allowing MPFs to adaptively control the sharpness of the phase transition and the degree of saturation of the soft sign indicator. Experiments on both synthetic and scanned thin-shell and thin-plate shapes demonstrate that MPFs preserve thin and layered structures more faithfully than recent SDF-based methods, while also enabling more robust training and more reliable surface extraction than UDF-based approaches. Check out \href{https://github.com/JIAYI-Scarlett/ICML2026-MPF}{MPFs-GitHub} for source code and test models.

cs.CV↗

arXiv2026

LycheeDecode: Accelerating Long-Context LLM Inference via Hybrid-Head Sparse Decoding

The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sharing undermines model performance by neglecting the functional diversity of attention heads. To address this, we propose LycheeDecode, an efficient decoding method centered on a fine-grained hybrid-head attention mechanism that employs a hardware-efficient top-k selection strategy. Specifically, the novel HardKuma-based mechanism partitions attention heads into a small subset of retrieval heads that dynamically identify crucial tokens and a majority of sparse heads that reuse them for efficient computation. Through extensive experiments on leading models like Llama3 and Qwen3 across diverse benchmarks for long-context understanding (e.g., LongBench, RULER) and complex reasoning (e.g., AIME24, OlympiadBench), we demonstrate that LycheeDecode achieves generative quality comparable to, and at times surpassing even the full-attention baseline. Crucially, this is accomplished with up to a 2.7x speedup at a 128K context length. By preserving the functional diversity of attention heads, our fine-grained strategy overcomes the performance bottlenecks of existing methods, providing a powerful and validated pathway to both efficient and high-quality long-context LLM inference.

cs.CL↗

arXiv2026

Sequential Operating Simulation of Solid State Transformer-Driven Next-Generation 800 VDC Data Center

Artificial-intelligence (AI) workloads are driving rapid growth in data-center electricity use and rack power density, increasing demand for power-delivery systems that are efficient and robust to fast load transients. Conventional uninterruptible power supply (UPS) based AC distribution chains involve multiple conversion stages and line-frequency transformers, which compound losses and are less compatible with dynamic AI power profiles. Although solid-state transformers (SSTs) and 800 VDC distribution architecture are widely discussed, implementable topology/control details, and long-horizon validation with realistic operating profiles remain limited. This paper develops an SST-driven 800 VDC architecture that converts 10 kV MVAC to an 800V LVDC bus using a three-phase H-bridge AC/DC stage cascaded with a dual-active-bridge (DAB) DC/DC stage. A coordinated closed-loop control scheme, combining rectifier voltage/current regulation and DAB phase-shift control, is designed to maintain DC-bus voltage stability. The proposed system is implemented on the real-time digital simulation (RTDS) platform and evaluated via sequential simulations using real-world day- and month-scale operating profiles of data centers, benchmarked against a UPS supply chain. Numerical studies demonstrate tight 800 VDC regulation, reduced input-side energy consumption compared with the UPS baseline, and satisfactory power-quality performance. A capacitance sensitivity test quantifies tradeoffs between DC-bus ripple and low-frequency input-power oscillations, yielding a practical capacitance range for design. Overall, the work provides a reproducible evaluation workflow and actionable guidance for next-generation AI data centers.

eess.SY↗

arXiv2025

Voronoi-Assisted Optimization for Diffusing Unsigned Distance Fields from Unoriented Points

Unsigned Distance Fields (UDFs) provide a flexible representation for 3D shapes with arbitrary topology, including open and closed surfaces, orientable and non-orientable geometries, and non-manifold structures. While recent neural approaches have shown promise in learning UDFs, they often suffer from numerical instability, high computational cost, and limited controllability. We present a lightweight, network-free method, Voronoi-Assisted Optimization for Diffusing (VAD), to compute UDFs directly from unoriented point clouds. Our approach begins by assigning bi-directional normals to input points, guided by two Voronoi-based geometric criteria encoded in an energy function for optimal alignment. The aligned normals are then diffused to form an approximate UDF gradient field, which is subsequently integrated to recover the final UDF. Experiments demonstrate that VAD robustly handles watertight and open surfaces, as well as complex non-manifold and non-orientable geometries, while remaining computationally efficient and stable.

cs.CV↗

arXiv2025

Is On-Policy Data always the Best Choice for Direct Preference Optimization-based LM Alignment?

The alignment of language models~(LMs) with human preferences is critical for building reliable AI systems. The problem is typically framed as optimizing an LM policy to maximize the expected reward that reflects human preferences. Recently, Direct Preference Optimization~(DPO) was proposed as a LM alignment method that directly optimize the policy from static preference data, and further improved by incorporating on-policy sampling~(i.e., preference candidates generated during the training loop) for better LM alignment. However, we show on-policy data is not always optimal, with systematic effectiveness difference emerging between static and on-policy preference candidates. For example, on-policy data can result in a $3\times$ effectiveness compared with static data for Llama-3, and a $0.4\times$ effectiveness for Zephyr. To explain the phenomenon, we propose the alignment stage assumption, which divides the alignment process into two distinct stages: the preference injection stage, which benefits from diverse data, and the preference fine-tuning stage, which favors high-quality data. Through theoretical and empirical analysis, we characterize these stages and propose an effective algorithm to identify the boundaries between them. We perform experiments on $5$ models~(Llama, Zephyr, Phi-2, Qwen, Pythia) and $2$ alignment methods~(DPO, SLiC-HF) to show the generalizability of alignment stage assumption and the effectiveness of the boundary measurement algorithm.

cs.AI↗

arXiv2025

MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction

Unsigned distance fields (UDFs) are widely used in 3D deep learning due to their ability to represent shapes with arbitrary topology. While prior work has largely focused on learning UDFs from point clouds or multi-view images, extracting meshes from UDFs remains challenging, as the learned fields rarely attain exact zero distances. A common workaround is to reconstruct signed distance fields (SDFs) locally from UDFs to enable surface extraction via Marching Cubes. However, this often introduces topological artifacts such as holes or spurious components. Moreover, local SDFs are inherently incapable of representing non-manifold geometry, leading to complete failure in such cases. To address this gap, we propose MIND (Material Interface from Non-manifold Distance fields), a novel algorithm for generating material interfaces directly from UDFs, enabling non-manifold mesh extraction from a global perspective. The core of our method lies in deriving a meaningful spatial partitioning from the UDF, where the target surface emerges as the interface between distinct regions. We begin by computing a two-signed local field to distinguish the two sides of manifold patches, and then extend this to a multi-labeled global field capable of separating all sides of a non-manifold structure. By combining this multi-labeled field with the input UDF, we construct material interfaces that support non-manifold mesh extraction via a multi-labeled Marching Cubes algorithm. Extensive experiments on UDFs generated from diverse data sources, including point cloud reconstruction, multi-view reconstruction, and medial axis transforms, demonstrate that our approach robustly handles complex non-manifold surfaces and significantly outperforms existing methods. The source code is available at https://github.com/jjjkkyz/MIND.

cs.CV↗

arXiv2024

From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $\alpha$-NeuS

Traditional 3D shape reconstruction techniques from multi-view images, such as structure from motion and multi-view stereo, face challenges in reconstructing transparent objects. Recent advances in neural radiance fields and its variants primarily address opaque or transparent objects, encountering difficulties to reconstruct both transparent and opaque objects simultaneously. This paper introduces $\alpha$-Neus -- an extension of NeuS -- that proves NeuS is unbiased for materials from fully transparent to fully opaque. We find that transparent and opaque surfaces align with the non-negative local minima and the zero iso-surface, respectively, in the learned distance field of NeuS. Traditional iso-surfacing extraction algorithms, such as marching cubes, which rely on fixed iso-values, are ill-suited for such data. We develop a method to extract the transparent and opaque surface simultaneously based on DCUDF. To validate our approach, we construct a benchmark that includes both real-world and synthetic scenes, demonstrating its practical utility and effectiveness. Our data and code are publicly available at https://github.com/728388808/alpha-NeuS.

cs.CV↗

arXiv2024

DCUDF2: Improving Efficiency and Accuracy in Extracting Zero Level Sets from Unsigned Distance Fields

Unsigned distance fields (UDFs) allow for the representation of models with complex topologies, but extracting accurate zero level sets from these fields poses significant challenges, particularly in preserving topological accuracy and capturing fine geometric details. To overcome these issues, we introduce DCUDF2, an enhancement over DCUDF--the current state-of-the-art method--for extracting zero level sets from UDFs. Our approach utilizes an accuracy-aware loss function, enhanced with self-adaptive weights, to improve geometric quality significantly. We also propose a topology correction strategy that reduces the dependence on hyper-parameter, increasing the robustness of our method. Furthermore, we develop new operations leveraging self-adaptive weights to boost runtime efficiency. Extensive experiments on surface extraction across diverse datasets demonstrate that DCUDF2 outperforms DCUDF and existing methods in both geometric fidelity and topological accuracy. We will make the source code publicly available.

cs.CV↗

arXiv2024

Diffusing Winding Gradients (DWG): A Parallel and Scalable Method for 3D Reconstruction from Unoriented Point Clouds

This paper presents a new method, Diffusing Winding Gradients (DWG), for reconstructing watertight 3D surfaces from unoriented point clouds. Our method exploits the alignment between the gradients of the generalized winding number (GWN) field and globally consistent normals to orient points effectively. Starting with an unoriented point cloud, DWG initially assigns a random normal to each point. It computes the corresponding GWN field and extract a level set whose iso-value is the average GWN values across all input points. The gradients of this level set are then utilized to update the point normals. This cycle of recomputing the GWN field and updating point normals is repeated until the GWN level sets stabilize and their gradients cease to change. Unlike conventional methods, our method does not rely on solving linear systems or optimizing objective functions, which simplifies its implementation and enhances its suitability for efficient parallel execution. Experimental results demonstrate that our method significantly outperforms existing methods in terms of runtime performance. For large-scale models with 10 to 20 million points, our CUDA implementation on an NVIDIA GTX 4090 GPU achieves speeds 30-120 times faster than iPSR, the leading sequential method, tested on a high-end PC with an Intel i9 CPU. Additionally, by employing a screened variant of GWN, DWG demonstrates enhanced robustness against noise and outliers, and proves effective for models with thin structures and real-world inputs with overlapping and misaligned scans. For source code and more details, visit our project webpage: https://dwgtech.github.io/.

cs.GR↗

arXiv2023

Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering

In this paper, we propose a new method, called DoubleCoverUDF, for extracting the zero level-set from unsigned distance fields (UDFs). DoubleCoverUDF takes a learned UDF and a user-specified parameter $r$ (a small positive real number) as input and extracts an iso-surface with an iso-value $r$ using the conventional marching cubes algorithm. We show that the computed iso-surface is the boundary of the $r$-offset volume of the target zero level-set $S$, which is an orientable manifold, regardless of the topology of $S$. Next, the algorithm computes a covering map to project the boundary mesh onto $S$, preserving the mesh's topology and avoiding folding. If $S$ is an orientable manifold surface, our algorithm separates the double-layered mesh into a single layer using a robust minimum-cut post-processing step. Otherwise, it keeps the double-layered mesh as the output. We validate our algorithm by reconstructing 3D surfaces of open models and demonstrate its efficacy and effectiveness on synthetic models and benchmark datasets. Our experimental results confirm that our method is robust and produces meshes with better quality in terms of both visual evaluation and quantitative measures than existing UDF-based methods. The source code is available at https://github.com/jjjkkyz/DCUDF.

cs.CV↗

arXiv2023

2S-UDF: A Novel Two-stage UDF Learning Method for Robust Non-watertight Model Reconstruction from Multi-view Images

Recently, building on the foundation of neural radiance field, various techniques have emerged to learn unsigned distance fields (UDF) to reconstruct 3D non-watertight models from multi-view images. Yet, a central challenge in UDF-based volume rendering is formulating a proper way to convert unsigned distance values into volume density, ensuring that the resulting weight function remains unbiased and sensitive to occlusions. Falling short on these requirements often results in incorrect topology or large reconstruction errors in resulting models. This paper addresses this challenge by presenting a novel two-stage algorithm, 2S-UDF, for learning a high-quality UDF from multi-view images. Initially, the method applies an easily trainable density function that, while slightly biased and transparent, aids in coarse reconstruction. The subsequent stage then refines the geometry and appearance of the object to achieve a high-quality reconstruction by directly adjusting the weight function used in volume rendering to ensure that it is unbiased and occlusion-aware. Decoupling density and weight in two stages makes our training stable and robust, distinguishing our technique from existing UDF learning approaches. Evaluations on the DeepFashion3D, DTU, and BlendedMVS datasets validate the robustness and effectiveness of our proposed approach. In both quantitative metrics and visual quality, the results indicate our superior performance over other UDF learning techniques in reconstructing 3D non-watertight models from multi-view images. Our code is available at https://bitbucket.org/jkdeng/2sudf/.

cs.CV↗

arXiv2022

Smart Power Supply for UAV Agility Enhancement Using Deep Neural Networks

Recently unmanned aerial vehicles (UAV) have been widely deployed in various real-world scenarios such as disaster rescue and package delivery. Many of these working environments are unstructured with uncertain and dynamic obstacles. UAV collision frequently happens. An UAV with high agility is highly desired to adjust its motions to adapt to these environmental dynamics. However, UAV agility is restricted by its battery power output; particularly, an UAV's power system cannot be aware of its actual power need in motion planning while the need is dynamically changing as the environment and UAV condition vary. It is difficult to accurately and timely align the power supply with power needs in motion plannings. This mismatching will lead to an insufficient power supply to an UAV and cause delayed motion adjustments, largely increasing the risk of collisions with obstacles and therefore undermine UAV agility. To improve UAV agility, a novel intelligent power solution, Agility-Enhanced Power Supply (AEPS), was developed to proactively prepare appropriate amount powers at the right timing to support motion planning with enhanced agility. This method builds a bridge between the physical power system and UAV planning. With agility-enhanced motion planning, the safety of UAV in complex working environment will be enhanced. To evaluate AEPS effectiveness, missions of "patrol missions for community security" with unexpected obstacles were adopted; the power supply is realized by hybrid integration of fuel cell, battery, and capacitor. The effectiveness of AEPS in improving UAV agility was validated by the successful and timely power supply, improved task success rate and system safety, and reduced mission duration.

cs.RO↗

arXiv2016

Attenuation of TeV $γ$-rays by the starlight photon field of the host galaxy

The absorption of TeV $γ$-ray photons produced in relativistic jets by surrounding soft photon fields is a long-standing problem of jet physics. In some cases the most likely emission site close to the central black hole is ruled out because of the high opacity caused by strong optical and infrared photon sources, such as the broad line region. Mostly neglected for jet modeling is the absorption of $γ$-rays in the starlight photon field of the host galaxy. Analyzing the absorption for arbitrary locations and observation angles of the $γ$-ray emission site within the host galaxy we find that the distance to the galaxy center, the observation angle, and the distribution of starlight in the galaxy are crucial for the amount of absorption. We derive the absorption value for a sample of $20$ TeV detected blazars with a redshift $z_r<0.2$. The absorption value of the $γ$-ray emission located in the galaxy center may be as high as $20\%$ with an average value of $6\%$. This is important in order to determine the intrinsic blazar parameters. We see no significant trends in our sample between the degree of absorption and host properties, such as starlight emissivity, galactic size, half-light radius, and redshift. While the uncertainty of the spectral properties of the extragalactic background light exceeds the effect of absorption by stellar light from the host galaxy in distant objects, the latter is a dominant effect in nearby sources. It may also be revealed in a differential comparison of sources with similar redshifts.

astro-ph.HE↗

arXiv2016

Particle diffusion and localized acceleration in inhomogeneous AGN jets - Part II: stochastic variation

We study the stochastic variation of blazar emission under a 2-D spatially resolved leptonic jet model we previously developed. Random events of particle acceleration and injection in small zones within the emission region are assumed to be responsible for flux variations. In addition to producing spectral energy distributions that describe the observed flux of Mrk 421, we further analyze the timing properties of the simulated light curves, such as the power spectral density (PSD) at different bands, flux-flux correlations, as well as the cross-correlation function between X-rays and TeV γ-rays. We find spectral breaks in the PSD at a timescale comparable to the dominant characteristic time scale in the system, which is usually the pre-defined decay time scale of an acceleration event. Cooling imposes a delay, and so PSDs taken at lower energy bands in each emission component (synchrotron or inverse Compton) generally break at longer timescales. The flux-flux correlation between X-rays and TeV γ-rays can be either quadratic or linear, depending on whether or not there are large variation of the injection into the particle acceleration process. When the relationship is quadratic, the TeV flares lag the X-ray flares, and the optical & GeV flares are large enough to be comparable to the ones in X-ray. When the relationship is linear, the lags are insignificant, and the optical & GeV flares are small.

astro-ph.HE↗

arXiv2015

Polarization Swings Reveal Magnetic Energy Dissipation in Blazars

The polarization signatures of the blazar emissions are known to be highly variable. In addition to small fluctuations of the polarization angle around a mean value, sometimes large (> 180^o) polarization angle swings are observed. We suggest that such p henomena can be interpreted as arising from light-travel-time effects within an underlying axisymmetric emission region. We present the first simultaneous fitting of the multi-wavelength spectrum, variability and time-dependent polarization features of a correlated optical and gamma-ray flaring event of the prominent blazar 3C279, which was accompanied by a drastic change of its polarization signatures. This unprecedented combination of spectral, variability, and polarization information in a coherent physical model allows us to place stringent constraints on the particle acceleration and magnetic-field topology in the relativistic jet of a blazar, strongly favoring a scenario in which magnetic energy dissipation is the primary driver of the flare event.

astro-ph.HE↗

arXiv2014

Particle diffusion and localized acceleration in inhomogeneous AGN jets - Part I: Steady-state spectra

We study the acceleration, transport, and emission of particles in relativistic jets. Localized stochastic particle acceleration, spatial diffusion, and synchrotron as well as synchrotron self-Compton emission are considered in a leptonic model. To account for inhomogeneity, we use a 2D axi-symmetric cylindrical geometry for both relativistic electrons and magnetic field. In this first phase of our work, we focus on steady-state spectra that develop from a time-dependent model. We demonstrate that small isolated acceleration region in a much larger emission volume are sufficient to accelerate particles to high energy. Diffusive escape from these small regions provides a natural explanation for the spectral form of the jet emission. The location of the acceleration regions within the jet is found to affect the cooling break of the spectrum in this diffusive model. Diffusion-caused energy-dependent inhomogeneity in the jets predicts that the SSC spectrum is harder than the synchrotron spectrum. There can also be a spectral hardening towards the high-energy section of the synchrotron spectrum, if particle escape is relatively slow. These two spectral hardening effects indicate that the jet inhomogeneity might be a natural explanation for the unexpected hard γ-ray spectra observed in some blazars.

astro-ph.HE↗