arXiv subjects

Kun Xu

Publications and source records attributed to Kun Xu.

At least 73 records · Page 4Linked to original sources

arXiv2026

A single-stage high-order compact gas-kinetic scheme in arbitrary Lagrangian-Eulerian formulation

This study presents the development of a compact gas-kinetic scheme using an arbitrary Lagrangian-Eulerian (ALE) formulation for structured meshes. Unlike the Eulerian formulation, the ALE approach effectively tracks flow discontinuities, such as shock waves and contact discontinuities. However, mesh motion alters the geometry and increases computational costs. To address this, two key strategies were introduced to reduce costs and enhance accuracy. The first strategy is to use the gas-kinetic scheme to construct a third-order gas-kinetic flux, rather than the Runge-Kutta method to achieve high-order time accuracy, which allows a single reconstruction and flux calculation per time step. This approach enables direct updates of both cell-averaged flow variables and their gradients using a time-accurate flux function, facilitating compact reconstruction. Second, the significant computational expense is spent on reconstruction, which requires recalculating the reconstruction matrix at each time step due to mesh changes. A simplified fourth-order compact reconstruction using a small matrix was used to mitigate this cost. The combination of fourth-order spatial reconstruction and third-order time-accurate flux evolution ensures both high resolution and computational efficiency in the ALE framework. The tests shows that the current reconstruction is 2.4x to 3.0x faster than the previous reconstruction. Additionally, a generalized ENO(GENO) method for handling discontinuities enhances the scheme's robustness. The numerical test cases, such as the Riemann problem, Sedov problem, Noh problem, and Saltzmann problem, demonstrated the robustness and accuracy of our method.

physics.comp-ph↗

arXiv2026

Wave Particle Turbulent Simulation of Spatially Developing Round Jets Using a Non Equilibrium Transport Model with a Mixing Length Characteristic Time Closure

In this paper, the wave-particle turbulent simulation (WPTS), a recently developed multiscale, non-equilibrium turbulence modeling approach, is coupled with a turbulence characteristic-time closure derived from Prandtl mixing-length hypothesis and applied to spatially developing round jets. In WPTS, fluid elements in strongly turbulent regions are represented by Lagrangian particles that travel a finite distance before interacting with the background flow field represented in a wave-like (Eulerian) form. This mechanism bears conceptual similarity to the discrete fluid parcels invoked in the Prandtl mixing-length picture. WPTS differs from conventional mixing-length-based turbulence models in two key respects. First, particle evolution follows a non-equilibrium transport mechanism, rather than the equilibrium assumptions typically embedded in eddy-viscosity closures. Second, WPTS advances the wave and particle components in a coupled manner, with the particle fraction governed primarily by the modeled turbulence characteristic time, enabling laminar and turbulent regimes to be represented within a unified framework. Because spatially developing jets provide a canonical test case with well-established similarity behavior, they are used here for evaluation. Specifically, this work (1) develops a mixing-length-based characteristic-time model tailored to jet flows and (2) incorporates it into WPTS to assess predictive performance. The resulting WPTS framework accurately reproduces the jet similarity solution and other characteristic features at Reynolds numbers of 5,000 and 20,000, demonstrating the promise of WPTS as a practical tool for turbulence modeling and simulation.

physics.flu-dyn↗

arXiv2026

WideSeek: Advancing Wide Research via Multi-Agent Scaling

Search intelligence is evolving from Deep Research to Wide Research, a paradigm essential for retrieving and synthesizing comprehensive information under complex constraints in parallel. However, progress in this field is impeded by the lack of dedicated benchmarks and optimization methodologies for search breadth. To address these challenges, we take a deep dive into Wide Research from two perspectives: Data Pipeline and Agent Optimization. First, we produce WideSeekBench, a General Broad Information Seeking (GBIS) benchmark constructed via a rigorous multi-phase data pipeline to ensure diversity across the target information volume, logical constraints, and domains. Second, we introduce WideSeek, a dynamic hierarchical multi-agent architecture that can autonomously fork parallel sub-agents based on task requirements. Furthermore, we design a unified training framework that linearizes multi-agent trajectories and optimizes the system using end-to-end RL. Experimental results demonstrate the effectiveness of WideSeek and multi-agent RL, highlighting that scaling the number of agents is a promising direction for advancing the Wide Research paradigm.

cs.CL↗

arXiv2026

LLM Latent Reasoning as Chain of Superposition

Latent reasoning offers a computation-efficient alternative to Chain-of-Thought but often suffers from performance degradation due to distributional misalignment and ambiguous chain definitions. Ideally, latent reasoning should function as a superposition of multiple reasoning paths. To realize this, we introduce Latent-SFT, a unified framework addressing challenges at three levels: token, chain, and learning. First, we define the Latent-Vocab to constrain hidden states within the pre-trained vocab-space. Second, we construct the Latent-Chain via Induction-Supervision Masking to ensure semantic compactness and sufficiency. Third, we employ Latent-Optim with stochastic Gumbel-Softmax to guide the model toward generalizable solutions. Empirical results demonstrate that Latent-SFT consistently outperforms explicit SFT across six mathematical benchmarks (e.g., GSM8k, AIME24) while achieving a 2.7x to 5.5x reduction in reasoning length. Analysis confirms that our method effectively captures a superposition of diverse reasoning trajectories rather than merely compressing a single path.

cs.CL↗

arXiv2026

The stability priority of spatial-temporal coupled compact element methods over decoupled compact element methods

With the increasing industrial demands, two families of high-order numerical schemes are widely used within the computational fluid dynamics community. One is the method of line, which relies on Runge-Kutta (RK) time-stepping applied to a semi-discrete, spatio-temporally decoupled formulation. The other is the family of Lax-Wendroff (LW) type method, which are inherently spatial-temporal coupled and are constructed within a multi-stage multi-derivative (MSMD) framework. This paper, for the first time, conducted a comparative Fourier stability analysis of RK and LW method to distinguish the dispersion and dissipation effects of numerical schemes respectively. Through rigorous theoretical derivation and consistent numerical validation, we draw the following conclusions: While explicit RK line methods are straightforward like Discontinuous Galerkin (DG) method and flux reconstruction (FR) method, they employ from a decoupling of spatial and temporal accuracy, thus discarding flow field evolution information and requiring small time steps. In contrast, spatial-temporal coupled compact methods, such as the gas-kinetic scheme (GKS) and the generalized Riemann problem (GRP) solver, utilize initial-value information from space far more effectively for time evolution. Even with just one additional order of spatial-temporal coupled information, they show better stability compared to RK methods. This provides new insights for CFD algorithm design, emphasizing the need for consistency between the dependence in the physical domain and that in numerical domain.

math-ph↗

arXiv2025

A Gas-Kinetic Scheme for Maxwell Equations

The Gas-Kinetic Scheme (GKS), widely used in computational fluid dynamics for simulating hypersonic and other complicated flow phenomena, is extended in this work to electromagnetic problems by solving Maxwell's equations. In contrast to the classical GKS formulation, the proposed scheme employs a discrete rather than a continuous velocity space. By evaluating a time-accurate numerical flux at cell interfaces, the proposed scheme attains second-order accuracy within a single step. Its kinetic formulation provides an inherently multidimensional framework, while the finite-volume formulation ensures straightforward extension to unstructured meshes. Through the incorporation of a collision process, the scheme exhibits lower numerical dissipation than classical flux-vector splitting (FVS) methods. Furthermore, the kinetic decomposition enables direct implementation of non-reflecting boundary conditions. The proposed scheme is validated against several benchmark problems and compared with established methods, including the Finite-Difference Time-Domain (FDTD) method and FVS. A lattice Boltzmann method (LBM) implementation is also included for comparative analysis. Finally, the technique is applied to simulate electromagnetic wave propagation in a realistic aircraft configuration, demonstrating its ability to model complex geometries.

math.NA↗

arXiv2025

The study of coherent Rayleigh-Brillouin scattering in multiple flow regimes using unified gas-kinetic scheme

Coherent Rayleigh-Brillouin scattering (CRBS) holds great promise for the characterization of gas properties and the investigation of gas kinetic processes. The CRBS spectrum exhibits a strong dependence on the Knudsen number (Kn), revealing its inherently multiscale nature. In the unified gas-kinetic scheme (UGKS), collisions are intrinsically coupled with free transport during flux construction, endowing the method with distinct multiscale capabilities. Specifically, the UGKS reduces to a Boltzmann solver when the relaxation time is greater than or equal to the time step, and to the gas-kinetic scheme (GKS)-a Navier-Stokes solver-when the relaxation time is much smaller than the time step, thereby accommodating flow regimes without constraints on the molecular mean free path or collision time. In this study, the UGKS is extended to simulate CRBS phenomena, with the governing equation formulated based on the BGK-Shakhov model. Detailed derivations are provided. To account for the additional perturbation source term, a second-order accurate numerical algorithm is developed using the Strang splitting method within the UGKS framework. The proposed model is validated against argon CRBS experiments, demonstrating excellent agreement. Building on this validated framework, the impact of incident signal intensity on CRBS spectra across a range of Knudsen numbers is systematically examined, accompanied by an in-depth analysis of the underlying physical mechanisms. This work broadens the applicability of CFD-based CRBS simulations and provides a reliable numerical foundation for exploring high-intensity, multiscale gas-kinetic phenomena in future research.

physics.flu-dyn↗

arXiv2025

UGKS and UGKWP Methods for Multiscale Simulation of Electrostatic Plasma in Quasineutral and Hydrodynamic Limits

This study extends the Unified Gas-Kinetic Scheme (UGKS) and the Unified Gas-Kinetic Wave-Particle (UGKWP) method for electrostatic plasma modeling, ensuring the correct asymptotic limits with respect to both the Debye length and the mean free path. By coupling collision and transport processes within the numerical flux, the proposed approach effectively removes the hydrodynamic-limit constraint associated with the mean free path. In addition, a reformulated Poisson equation, coupled with the macroscopic moment equations, is introduced to overcome the inefficiency of the standard Poisson formulation in the quasineutral regime. The accuracy and asymptotic consistency of the proposed schemes are verified through several benchmark tests, including linear and nonlinear Landau damping and the bump-on-tail instability. The results demonstrate that the methods robustly capture plasma dynamics across hydrodynamic and quasineutral regimes, without resolution constraints imposed by either the Debye length or the mean free path.

physics.comp-ph↗

arXiv2025

Abundance and phase-space distribution of subhalos in cosmological N-body simulations: testing numerical convergence and correction methods

Subhalos play a crucial role in accurately modeling galaxy formation and galaxy-based cosmological probes within the highly nonlinear, virialized regime. However, numerical convergence of subhalo evolution is difficult to achieve, especially in the inner regions of host halos where tidal forces are strongest. I investigate the numerical convergence and correction methods for the abundance, spatial, and velocity distributions of subhalos using two $6144^3$-particle cosmological N-body simulations with different mass resolutions -- Jiutian-300 ($1.0 \times 10^{7}\,h^{-1}M_{\odot}$) and Jiutian-1G ($3.7 \times 10^{8}\,h^{-1}M_{\odot}$) -- with subhalos identified by HBT+. My study shows that the Surviving subhalo Peak Mass Function (SPMF) converges only for subhalos with $m_{\mathrm{peak}}$ above $5000$ particles but can be accurately recovered by including orphan subhalos that survive according to the merger timescale model of Jiang et al., which outperforms other models. Including orphan subhalos also enables recovery of the real-space spatial and velocity distributions to $5$--$10\%$ accuracy down to scales of $0.1$--$0.2\,h^{-1}\mathrm{Mpc}$. The remaining differences are likely due to cosmic variance and finite-box effects in the smaller Jiutian-300 simulation. Convergence below $0.1\,h^{-1}\mathrm{Mpc}$ remains challenging and requires more sophisticated modeling of orphan subhalos. I further highlight that redshift-space multipoles are more difficult to recover even at larger scales because unreliable small-scale pairs at $r_{\mathrm{p}} < 0.1\,h^{-1}\mathrm{Mpc}$ in real space affect scales of tens of $\mathrm{Mpc}$ in redshift space due to elongated Fingers-of-God effects. Therefore, for redshift-space statistics, I recommend using modified or alternative measures that reduce sensitivity to small projected separations in subhalo-based studies.

astro-ph.CO↗

arXiv2025

Generating Attribute-Aware Human Motions from Textual Prompt

Text-driven human motion generation has recently attracted considerable attention, allowing models to generate human motions based on textual descriptions. However, current methods neglect the influence of human attributes-such as age, gender, weight, and height-which are key factors shaping human motion patterns. This work represents a pilot exploration for bridging this gap. We conceptualize each motion as comprising both attribute information and action semantics, where textual descriptions align exclusively with action semantics. To achieve this, a new framework inspired by Structural Causal Models is proposed to decouple action semantics from human attributes, enabling text-to-semantics prediction and attribute-controlled generation. The resulting model is capable of generating attribute-aware motion aligned with the user's text and attribute inputs. For evaluation, we introduce a comprehensive dataset containing attribute annotations for text-motion pairs, setting the first benchmark for attribute-aware motion generation. Extensive experiments validate our model's effectiveness.

cs.CV↗

arXiv2025

Single-Ion Sensing in Liquid Using Fluorescent h-BN Point Defects

Understanding the chemical state of individual ions in solutions is crucial for advancing knowledge of complex systems. However, sensing systems at the single-ion level in liquid environments remains a significant challenge. A strategy is introduced that leverages the optical emission properties of point defects in hexagonal boron nitride (h-BN) as single ion sensors. The interaction of optically active h-BN defects with ions in solution leads to distinct spectral shifts, enabling precise visualization and analyzing of individual ions. Using Li+ ions in organic electrolytes as a model, spectral shifts exceeding 10 nm were observed upon ion addition. Application of an external electric field further enhanced these shifts to over 40 nm, enabling real-time monitoring of electrical field induced local perturbations of Li+ ions. Through this approach, individual point defects were shown to spectroscopically distinguish ions of varying charges (e.g., Na+, Mg2+, and Al3+) based on their local electrical field, each producing a distinct spectral shift. This platform allows direct sensing of ions and their chemical states in liquid environments, providing insights into subtle interfacial changes at the single-ion level, with measurable spectral shifts detectable at millisecond temporal resolution and at concentrations down to 10 micromolar range. This capability presents potential applications in various fields involving ions in liquids that include battery technology and environmental science.

cond-mat.mes-hall↗

arXiv2025

Defect Engineered Hexagonal-Boron Nitride Enables Ionic Conduction for Lithium Metal Batteries

The practical implementation of lithium-metal anodes has been hindered by uncontrollable dendrite formation and interfacial instability. This study presents a defect-engineering approach of a chemically stable and electrically insulating interfacial layer of hexagonal boron nitride (h-BN) that markedly enhances ionic conductivity through argon ion irradiation. Initially, the electrochemical performance from commercially available, large-area chemical vapor deposition (CVD)-grown h-BN films with industrial-scale argon ion implantation motivated our subsequent detailed investigations using lab-scale exfoliated single-crystal h-BN flakes. Integration of these exfoliated flakes into a hybrid microfluidic-microelectronic chip provided direct evidence that controlled vacancy defects transform h-BN into an efficient lithium-ion conductor while preserving its intrinsic electrical insulation. Experimental validation confirmed improved lithium-metal anode stability, achieving dendrite-free cycling with Li plating/stripping Coulombic efficiencies exceeding 99.5% about 1000 cycles. Further assemble of irradiated h-BN in lithium-sulfur batteries effectively mitigates the polysulfide shuttle effect, sustaining over 97% specific capacity around 300 cycles. These results establish a robust, scalable interface-engineering route for next-generation lithium-metal batteries that combine high ionic transport with excellent electrical insulation.

cond-mat.mtrl-sci↗

arXiv2025

Following the Autoregressive Nature of LLM Embeddings via Compression and Alignment

A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs' pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.

cs.CL↗

arXiv2025

Ultra-chaotic property of Navier-Stokes turbulence

A chaotic system is called ultra-chaos when its statistics have sensitivity dependence on initial condition and/or other small disturbances. In this paper, using two-dimensional turbulent Kolmogorov flow as an example, we illustrate that tiny variation of initial condition of Navier-Stokes equations can lead to huge differences not only in spatiotemporal trajectory but also in flow symmetry and its statistics. Here, in order to avoid the influence of artificial numerical noise, we apply ``clean numerical simulation'' (CNS) which can guarantee that the numerical noise can be reduced to such a desired low level that they are negligible in a time interval long enough for calculating statistics. This discovery highly suggests that the Navier-Stokes turbulence (i.e. turbulence governed by the Navier-Stokes equations) might be an ultra-chaos, say, small disturbances must be considered even from viewpoint of statistics. This however leads to a paradox in logic, since small disturbances, which are unavoidable in practice, are unfortunately neglected by the Navier-Stokes turbulence. Some fundamental characteristics of turbulence model are discussed and suggested in general meanings.

nlin.CD↗

arXiv2025

On the Diminishing Returns of Complex Robust RAG Training in the Era of Powerful LLMs

Retrieval-augmented generation (RAG) systems traditionally employ sophisticated training strategies to enhance robustness against retrieval noise. In this work, we investigate a critical question: does the benefit of these complex robust training methods diminish as language models become more powerful? Through systematic evaluation across multiple model scales and question-answering datasets, our analysis reveals a consistent trend: \emph{the marginal robustness benefit of sophisticated training strategies decreases substantially as model capacity increases.} While smaller models show significant performance improvements from complex document selection and adversarial objectives, more capable models achieve comparable or even superior performance with simpler training approaches. Further investigation demonstrates that stronger models naturally exhibit better confidence calibration, cross-dataset generalization capability, and more effective attention patterns, even under simple training regimes. These findings suggest that as foundation models evolve, the engineering effort invested in complex robust training may yield diminishing returns, indicating that simplified RAG pipelines could suffice for powerful models while maintaining competitive performance.

cs.CL↗

arXiv2025

A unified gas-kinetic framework from Boltzmann to Navier-Stokes scales

Neither molecular kinetics nor continuum fluid dynamics alone is adequate to describe multiscale gas flows across different regimes. Bridging these regimes within a single self-consistent framework has long been a central challenge in fluid mechanics. We propose a unified gas kinetic framework that classifies molecules by their collision histories over an observation timescale. This formulation recovers the Boltzmann and Navier-Stokes equations as limiting cases, providing a transparent connection between kinetic and hydrodynamic descriptions. Beyond practical advantages for multiscale modeling, this framework offers a new perspective on Hilbert's sixth problem by linking microscopic dynamics to continuum mechanics through a tunable observational scale.

physics.flu-dyn↗

arXiv2025

InstanceBEV: Unifying Instance and BEV Representation for 3D Panoptic Segmentation

BEV-based 3D perception has emerged as a focal point of research in end-to-end autonomous driving. However, existing BEV approaches encounter significant challenges due to the large feature space, complicating efficient modeling and hindering effective integration of global attention mechanisms. We propose a novel modeling strategy, called InstanceBEV, that synergistically combines the strengths of both map-centric approaches and object-centric approaches. Our method effectively extracts instance-level features within the BEV features, facilitating the implementation of global attention modeling in a highly compressed feature space, thereby addressing the efficiency challenges inherent in map-centric global modeling. Furthermore, our approach enables effective multi-task learning without introducing additional module. We validate the efficiency and accuracy of the proposed model through predicting occupancy, achieving 3D occupancy panoptic segmentation by combining instance information. Experimental results on the OCC3D-nuScenes dataset demonstrate that InstanceBEV, utilizing only 8 frames, achieves a RayPQ of 15.3 and a RayIoU of 38.2. This surpasses SparseOcc's RayPQ by 9.3% and RayIoU by 10.7%, showcasing the effectiveness of multi-task synergy.

cs.CV↗

arXiv2025

Electromagnetic Flow Control in Hypersonic Rarefied Environment

The Unified Gas-Kinetic Wave-Particle (UGKWP) method, developed for multiscale simulation of partially ionized plasmas, has been extended to unstructured meshes, enabling the modeling of electromagnetic flows around a hemisphere across near-continuum to rarefied regimes. To the best of our knowledge, this work represents the first application of a multiscale plasma solver to this problem. In our approach, neutrals, ions, and electrons are treated as distinct species, with electrons modeled beyond the conventional fluid approximation. The numerical implementation is validated through comparison with reference solutions for neutral hypersonic flow around a sphere, as well as benchmarking against experimental data for a Mach 4.75 pre-ionized argon flow. In both cases, the UGKWP results show good agreement with the reference and experimental data. The findings reveal that rarefied effects play a significant role in the prediction of electromagnetic flow control, underscoring the necessity of multiscale modeling in plasma flow applications.

physics.plasm-ph↗