arXiv subjects

Xinlei Zhang

Publications and source records attributed to Xinlei Zhang.

13 recordsLinked to original sources

arXiv2026

G0.5: One Autoregressive Stream for Robot Reasoning and Action

The prevailing recipe for Vision-Language-Action (VLA) models couples a pretrained VLM with a separately trained flow-matching action expert. This makes the VLM a context encoder rather than a decision-maker. We introduce G0.5, a pretrained autoregressive VLA in which a single transformer decoder emits reasoning and action tokens under a single objective. Three components make this tractable at foundation-model scale: a learnable cross-embodiment action tokenizer that maps heterogeneous robot actions into a shared vocabulary; a native chain-of-thought stream interleaving task decomposition, object grounding, and action hints with action tokens; and a visual memory module that injects multi-second history through the vision encoder. Because reasoning and action share a single set of weights, the pretrained VLM's capabilities carry over to physical behavior: the model follows instructions closely, and prompts directly steer action granularity, task horizon, and out-of-distribution scene handling without further training. Pretrained on a large collection of robot datasets together with VQA samples, G0.5 surpasses state-of-the-art models across 7 independent regimes: real-world fine-tuning on R1lite and R1pro robots (76.7\% vs.\ 53.3\% for $π_{0.5}$ and 24.4\% for GR00T-N1.7), the 2025 BEHAVIOR Challenge on 50 long-horizon household mobile manipulation tasks using a generalist policy (31.4\% vs.\ 26.3\% for $π_{0.5}$ and 26.1\% for the challenge winner), DROID post-training followed by zero-shot transfer to an unseen environment and objects (82.5\%), a language-following Pick-and-Place benchmark, LIBERO (98.9\%), RoboTwin 2.0 (93.3\%), and SimplerEnv-Bridge (87.3\%).

cs.RO↗

arXiv2026

Label-Noise Resistant Learning via Optimal Brain Damage Masking

Noisy labels are inevitable in real-world multimedia applications. Due to the strong memorization capacity of deep neural networks, these noisy labels cause significant performance degradation. Existing noise-robust methods have mainly focused on robust loss functions and sample selection strategies, with comparatively limited exploration of dynamic architectural adaptation. In this paper, we rethink the role of classifier connectivity under label noise. Intuitively, performance degradation stems from the backpropagation of noisy gradients. Since the final classifier layer acts as the primary gateway for this error propagation, selectively discarding redundant connections can restrict the backpropagation pathways of noisy gradients. Consequently, to identify redundant connections, we leverage the seminal Optimal Brain Damage (OBD) theory from model compression, which posits that parameters causing negligible loss perturbation can be removed. Guided by this principle, we show that masking low-activation edges limits the estimated loss perturbation to preserve the model's fitting capacity, while tightening a derived upper bound on noise-induced gradient error. To bridge this theoretical insight with practical training, we propose a novel Selective Edge Masking (SEM) mechanism for the fully connected (FC) layer to enhance noise robustness. It can adaptively retain critical edges for information propagation while suppressing gradient errors caused by noisy labels. As a plug-and-play component, SEM can be seamlessly integrated into various noise-robust methods. Additionally, we validate SEM's applicability by applying it to the newly emerged Kolmogorov-Arnold Network (KAN) employed as a classifier. Extensive evaluations on synthetic and real-world benchmarks demonstrate that our approach achieves state-of-the-art performance.

cs.LG↗

arXiv2025

Predicting Credit Spreads and Ratings with Machine Learning: The Role of Non-Financial Data

We build a 167-indicator comprehensive credit risk indicator set, integrating macro, corporate financial, bond-specific indicators, and for the first time, 30 large-scale corporate non-financial indicators. We use seven machine learning models to construct a bond credit spread prediction model, test their spread predictive power and economic mechanisms, and verify their credit rating prediction effectiveness. Results show these models outperform Chinese credit rating agencies in explaining credit spreads. Specially, adding non-financial indicators more than doubles their out-of-sample performance vs. traditional feature-driven models. Mechanism analysis finds non-financial indicators far more important than traditional ones (macro-level, financial, bond features)-seven of the top 10 are non-financial (e.g., corporate governance, property rights nature, information disclosure evaluation), the most stable predictors. Models identify high-risk traits (deteriorating operations, short-term debt, higher financing constraints) via these indicators for spread prediction and risk identification. Finally, we pioneer a credit rating model using predicted spreads (predicted implied rating model), with full/sub-industry models achieving over 75% accuracy, recall, F1. This paper provides valuable guidance for bond default early warning, credit rating, and financial stability.

econ.GN↗

arXiv2025

Efficient Pre-Training of LLMs via Topology-Aware Communication Alignment on More Than 9600 GPUs

The scaling law for large language models (LLMs) depicts that the path towards machine intelligence necessitates training at large scale. Thus, companies continuously build large-scale GPU clusters, and launch training jobs that span over thousands of computing nodes. However, LLM pre-training presents unique challenges due to its complex communication patterns, where GPUs exchange data in sparse yet high-volume bursts within specific groups. Inefficient resource scheduling exacerbates bandwidth contention, leading to suboptimal training performance. This paper presents Arnold, a scheduling system summarizing our experience to effectively align LLM communication patterns with data center topology at scale. An in-depth characteristic study is performed to identify the impact of physical network topology to LLM pre-training jobs. Based on the insights, we develop a scheduling algorithm to effectively align communication patterns with the physical network topology in modern data centers. Through simulation experiments, we show the effectiveness of our algorithm in reducing the maximum spread of communication groups by up to $1.67$x. In production training, our scheduling system improves the end-to-end performance by $10.6\%$ when training with more than $9600$ GPUs, a significant improvement for our training pipeline.

cs.DC↗

arXiv2025

Mycroft: Tracing Dependencies in Collective Communication Towards Reliable LLM Training

Reliability is essential for ensuring efficiency in LLM training. However, many real-world reliability issues remain difficult to resolve, resulting in wasted resources and degraded model performance. Unfortunately, today's collective communication libraries operate as black boxes, hiding critical information needed for effective root cause analysis. We propose Mycroft, a lightweight distributed tracing and root cause analysis system designed to address previously hidden reliability issues in collective communication. Mycroft's key idea is to trace collective communication states and leverage internal control and data dependencies to resolve reliability problems in LLM training. Mycroft has been deployed at ByteDance for over six months to debug collective communication related issues at runtime. It detected anomalies within 15 seconds in 90% of cases and identified the root cause within 20 seconds in 60% of cases. We also conducted extensive fault injection experiments to demonstrate Mycroft's capability and efficiency.

cs.DC↗

arXiv2025

MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism

Mixture-of-Experts (MoE) showcases tremendous potential to scale large language models (LLMs) with enhanced performance and reduced computational complexity. However, its sparsely activated architecture shifts feed-forward networks (FFNs) from being compute-intensive to memory-intensive during inference, leading to substantially lower GPU utilization and increased operational costs. We present MegaScale-Infer, an efficient and cost-effective system for serving large-scale MoE models. MegaScale-Infer disaggregates attention and FFN modules within each model layer, enabling independent scaling, tailored parallelism strategies, and heterogeneous deployment for both modules. To fully exploit disaggregation in the presence of MoE's sparsity, MegaScale-Infer introduces ping-pong pipeline parallelism, which partitions a request batch into micro-batches and shuttles them between attention and FFNs for inference. Combined with distinct model parallelism for each module, MegaScale-Infer effectively hides communication overhead and maximizes GPU utilization. To adapt to disaggregated attention and FFN modules and minimize data transmission overhead (e.g., token dispatch), MegaScale-Infer provides a high-performance M2N communication library that eliminates unnecessary GPU-to-CPU data copies, group initialization overhead, and GPU synchronization. Experimental results indicate that MegaScale-Infer achieves up to 1.90x higher per-GPU throughput than state-of-the-art solutions.

cs.DC↗

arXiv2025

Performance Evaluation of Large Language Models in Statistical Programming

The programming capabilities of large language models (LLMs) have revolutionized automatic code generation and opened new avenues for automatic statistical analysis. However, the validity and quality of these generated codes need to be systematically evaluated before they can be widely adopted. Despite their growing prominence, a comprehensive evaluation of statistical code generated by LLMs remains scarce in the literature. In this paper, we assess the performance of LLMs, including two versions of ChatGPT and one version of Llama, in the domain of SAS programming for statistical analysis. Our study utilizes a set of statistical analysis tasks encompassing diverse statistical topics and datasets. Each task includes a problem description, dataset information, and human-verified SAS code. We conduct a comprehensive assessment of the quality of SAS code generated by LLMs through human expert evaluation based on correctness, effectiveness, readability, executability, and the accuracy of output results. The analysis of rating scores reveals that while LLMs demonstrate usefulness in generating syntactically correct code, they struggle with tasks requiring deep domain understanding and may produce redundant or incorrect results. This study offers valuable insights into the capabilities and limitations of LLMs in statistical programming, providing guidance for future advancements in AI-assisted coding systems for statistical analysis.

stat.AP↗

arXiv2024

Boost UAV-based Ojbect Detection via Scale-Invariant Feature Disentanglement and Adversarial Learning

Detecting objects from Unmanned Aerial Vehicles (UAV) is often hindered by a large number of small objects, resulting in low detection accuracy. To address this issue, mainstream approaches typically utilize multi-stage inferences. Despite their remarkable detecting accuracies, real-time efficiency is sacrificed, making them less practical to handle real applications. To this end, we propose to improve the single-stage inference accuracy through learning scale-invariant features. Specifically, a Scale-Invariant Feature Disentangling module is designed to disentangle scale-related and scale-invariant features. Then an Adversarial Feature Learning scheme is employed to enhance disentanglement. Finally, scale-invariant features are leveraged for robust UAV-based object detection. Furthermore, we construct a multi-modal UAV object detection dataset, State-Air, which incorporates annotated UAV state parameters. We apply our approach to three lightweight detection frameworks on two benchmark datasets. Extensive experiments demonstrate that our approach can effectively improve model accuracy and achieve state-of-the-art (SoTA) performance on two datasets. Our code and dataset will be publicly available once the paper is accepted.

cs.CV↗

arXiv2021

Bayesian optimization for the spanwise oscillation of a gliding flat-plate

The kinematics of a gliding flat-plate with spanwise oscillation has been optimized to enhance the power efficiency by using Bayesian optimization method, in which the portfolio allocation framework consists of a Gaussian process probabilistic surrogate and a hybrid acquisition strategy. We tune three types of acquisition function in the optimization framework and assign three different balance parameters to each acquisition function. The design variables are set as the dimensionless oscillating amplitude and reduced frequency of the spanwise oscillation. The object function is to maximize the power factor to support a unit weight. The optimization results in a maximal power factor of 1.65 when the dimensionless oscillating amplitude and reduced frequency vary from 0 to 1. The features of the probabilistic response surface are also examined. There exists an optimal reduced frequency for the power efficiency at the oscillating amplitudes above 0.40. In addition, the higher power efficiency may be obtained by increasing the amplitude beyond 1.00.

physics.flu-dyn↗

arXiv2020

Effects of space sizes on the dispersion of cough-generated droplets from a walking person

The dispersion of viral droplets plays a key role in the transmission of COVID-19. In this work, we analyze the dispersion of cough-generated droplets in the wake of a walking person for different space sizes. The air flow is simulated by solving the Reynolds-Averaged Navier-Stokes equations, and the droplets are modelled as passive Lagrangian particles. Simulation results show that the cloud of droplets locates around and below the waist height of the manikin after two seconds from coughing, which indicates that kids walking behind an infectious patient are exposed to higher transmission risk than adults. More importantly, two distinct droplet dispersion modes occupying significantly different contamination regions are discovered. A slight change of space size is found being able to trigger the transition of dispersion modes even though the flow patterns are still similar. This shows the importance of accurately simulating the air flow in predicting the dispersion of viral droplets and implies the necessity to set different safe-distancing guidelines for different environments.

physics.flu-dyn↗

arXiv2020

The motion of respiratory droplets produced by coughing

Coronavirus disease 2019 (COVID-19) has become a global pandemic infectious respiratory disease with high mortality and infectiousness. This paper investigates respiratory droplet transmission, which is critical to understanding, modeling and controlling epidemics. In the present work, we implemented flow visualization, particle image velocimetry (PIV) and particle shadow tracking velocimetry (PSTV) to measure the velocity of the airflow and droplets involved in coughing and then constructed a physical model considering the evaporation effect to predict the motion of droplets under different weather conditions. The experimental results indicate that the convection velocity of cough airflow presents the relationship $t^{-0.7}$ with time; hence, the distance from the cougher increases by $t^{0.3}$ in the range of our measurement domain. Substituting these experimental results into the physical model reveals that the small droplets (initial diameter $D \leq$ 100 $μ$m) evaporate to droplet nuclei and that the large droplets with $D \geq$ 500 $μ$m and initial velocity $u_0 \geq$ 5 m/s travel more than 2 m. Winter conditions of low temperature and high relative humidity can cause more droplets to settle to the ground, which may be a possible driver of a second pandemic wave in the autumn and winter seasons.

physics.flu-dyn↗

arXiv2020

Evaluation of ensemble methods for quantifying uncertainties in steady-state CFD applications with small ensemble sizes

Bayesian uncertainty quantification (UQ) is of interest to industry and academia as it provides a framework for quantifying and reducing the uncertainty in computational models by incorporating available data. For systems with very high computational costs, for instance, the computational fluid dynamics (CFD) problem, the conventional, exact Bayesian approach such as Markov chain Monte Carlo is intractable. To this end, the ensemble-based Bayesian methods have been used for CFD applications. However, their applicability for UQ has not been fully analyzed and understood thus far. Here, we evaluate the performance of three widely used iterative ensemble-based data assimilation methods, namely ensemble Kalman filter, ensemble randomized maximum likelihood method, and ensemble Kalman filter with multiple data assimilation for UQ problems. We present the derivations of the three ensemble methods from an optimization viewpoint. Further, a scalar case is used to demonstrate the performance of the three different approaches with emphasis on the effects of small ensemble sizes. Finally, we assess the three ensemble methods for quantifying uncertainties in steady-state CFD problems involving turbulent mean flows. Specifically, the Reynolds averaged Navier--Stokes (RANS) equation is considered the forward model, and the uncertainties in the propagated velocity are quantified and reduced by incorporating observation data. The results show that the ensemble methods cannot accurately capture the true posterior distribution, but they can provide a good estimation of the uncertainties even when very limited ensemble sizes are used. Based on the overall performance and efficiency from the comparison, the ensemble randomized maximum likelihood method is identified as the best choice of approximate Bayesian UQ approach~among the three ensemble methods evaluated here.

physics.comp-ph↗

arXiv2019

Enforcing Boundary Conditions on Physical Fields in Bayesian Inversion

Inverse problems in computational mechanics consist of inferring physical fields that are latent in the model describing some observable fields. For instance, an inverse problem of interest is inferring the Reynolds stress field in the Navier--Stokes equations describing mean fluid velocity and pressure. The physical nature of the latent fields means they have their own set of physical constraints, including boundary conditions. The inherent ill-posedness of inverse problems, however, means that there exist many possible latent fields that do not satisfy their physical constraints while still resulting in a satisfactory agreement in the observation space. These physical constraints must therefore be enforced through the problem formulation. So far there has been no general approach to enforce boundary conditions on latent fields in inverse problems in computational mechanics, with these constraints often simply ignored. In this work we demonstrate how to enforce boundary conditions in Bayesian inversion problems by choice of the statistical model for the latent fields. Specifically, this is done by modifying the covariance kernel to guarantee that all realizations satisfy known values or derivatives at the boundary. As a test case the problem of inferring the eddy viscosity in the Reynolds-averaged Navier--Stokes equations is considered. The results show that enforcing these constraints results in similar improvements in the output fields but with latent fields that behave as expected at the boundaries.

physics.comp-ph↗