arXiv subjects

Zeyu Li

Publications and source records attributed to Zeyu Li.

At least 55 records · Page 3Linked to original sources

arXiv2025

Bridging Semantic Logic Gaps: A Cognition Inspired Multimodal Boundary Preserving Network for Image Manipulation Localization

The existing image manipulation localization (IML) models mainly relies on visual cues, but ignores the semantic logical relationships between content features. In fact, the content semantics conveyed by real images often conform to human cognitive laws. However, image manipulation technology usually destroys the internal relationship between content features, thus leaving semantic clues for IML. In this paper, we propose a cognition inspired multimodal boundary preserving network (CMB-Net). Specifically, CMB-Net utilizes large language models (LLMs) to analyze manipulated regions within images and generate prompt-based textual information to compensate for the lack of semantic relationships in the visual information. Considering that the erroneous texts induced by hallucination from LLMs will damage the accuracy of IML, we propose an image-text central ambiguity module (ITCAM). It assigns weights to the text features by quantifying the ambiguity between text and image features, thereby ensuring the beneficial impact of textual information. We also propose an image-text interaction module (ITIM) that aligns visual and text features using a correlation matrix for fine-grained interaction. Finally, inspired by invertible neural networks, we propose a restoration edge decoder (RED) that mutually generates input and output features to preserve boundary information in manipulated regions without loss. Extensive experiments show that CMB-Net outperforms most existing IML models. Our code is available on https://github.com/vpsg-research/CMB-Net.

cs.CV↗

arXiv2025

Impact of shape coexistence on the symmetric to asymmetric fission mode transition in Th isotopes

We study the evolution of fission modes along the Th isotopic chain using a microscopic framework combining the time-dependent generator coordinate method and finite-temperature covariant density functional theory. Theoretical fission fragment charge distributions agree well with experiments, and reveal a rapid symmetric-to-asymmetric transition from $A=222$ to 234. By analyzing the collective potential energy surfaces and time evolution of collective probability density distributions, we demonstrate that this fission mode transition is strongly correlated with the rapidly deepening asymmetric fission valley $-$ a phenomenon driven by the reduction of deformation energies of both the heavy and light fragments formed in the asymmetric fission valley. Further analysis attributes the decrease of light-fragment deformation energies to the onset of a coexisting large-deformed minimum in neutron-rich Kr and Sr isotopes (dominated isotopes for light asymmetric peak), which arises from a deformed proton $Z=38$ shell closure near $β_2\approx0.46$. Notably, we identify, for the first time, the pivotal role of the light fragment and its shape coexistence structure on the fission mode transition in Th isotopes in a fully microscopic framework.

nucl-th↗

arXiv2025

Dissecting the NVIDIA Hopper Architecture through Microbenchmarking and Multiple Level Analysis

This study presents a comprehensive multi-level analysis of the NVIDIA Hopper GPU architecture, focusing on its performance characteristics and novel features. We benchmark Hopper's memory subsystem, highlighting improvements in the L2 partitioned cache and global memory access compared to Ampere and Ada Lovelace. The evaluation of Hopper's fourth-generation tensor cores reveals the benefits of FP8 precision and asynchronous wgmma instructions for matrix operations. Additionally, we investigate the performance of DPX instructions for dynamic programming, distributed shared memory (DSM) for inter-SM communication, and the Tensor Memory Accelerator (TMA) for asynchronous data movement. Through multi-level evaluation, we discover that the Hopper architecture demonstrates significant acceleration potential in real-world applications. For instance, the asynchronous programming model supported by TMA achieves a 1.5x speedup in matrix multiplication, FP8 delivers nearly double the performance of FP16, and DPX instructions accelerate a computational biology algorithm by at least 4.75x. Our findings provide actionable insights for optimizing compute-intensive workloads, from AI training to bioinformatics, on Hopper GPUs.

cs.DC↗

arXiv2025

Engineering topological phase transitions via sliding ferroelectricity in MBi2Te4 (M = Ge, Sn, Pb) bilayers

Materials combining electrically switchable ferroelectricity and tunable topological states hold significant promise for advancing both foundamental quantum phenomena and innovative device architectures. Here, we employ first-principles calculations to systematically investigate the sliding ferroelectricity-mediated topological transitions in bilayer MBi2Te4 (M = Ge, Sn, Pb). By strategically engineering interlayer sliding configurations with oppositely polarized states, we demonstrate reversible band inversion accompanied by topological phase transitions. The calculated spin-orbit-coupled bandgaps reach 31 meV (GeBi2Te4), 36 meV (SnBi2Te4), and 35 meV (PbBi2Te4), thereby enabling room-temperature observation of the quantum spin Hall effect. Crucially, these systems exhibit substantial out-of-plane ferroelectric polarization magnitudes of 0.571-0.623 pC/m, with PbBi2Te4 showing the maximum polarization (0.623 pC/m). The topological nontriviality is unambiguously confirmed by two independent signatures: (i) the computed z2 topological invariant, and (ii) the emergence of gapless helical edge states spanning the bulk insulating gap. This synergy arises from the unique sliding-induced charge redistribution mechanism, which simultaneously modulates Berry curvature and breaks in-plane inversion symmetry without disrupting out-of-plane polarization stability. The co-engineering of non-volatile ferroelectric switching and topologically protected conduction channels in MBi2Te4 bilayers establishes a material paradigm for designing reconfigurable quantum devices, where electronic topology can be electrically controlled via polarization reversal. Our results provide critical insights into manipulating correlated quantum states in van der Waals ferroelectrics for multifunctional nanoelectronics.

cond-mat.mtrl-sci↗

arXiv2025

BurstGPT: A Real-world Workload Dataset to Optimize LLM Serving Systems

Serving systems for Large Language Models (LLMs) are often optimized to improve quality of service (QoS) and throughput. However, due to the lack of open-source LLM serving workloads, these systems are frequently evaluated under unrealistic workload assumptions. Consequently, performance may degrade when systems are deployed in real-world scenarios. This work presents BurstGPT, an LLM serving workload with 10.31 million traces from regional Azure OpenAI GPT services over 213 days. BurstGPT captures LLM serving characteristics from user, model and system perspectives: (1) User request concurrency: burstiness variations of requests in Azure OpenAI GPT services, revealing diversified concurrency patterns in different services and model types. (2) User conversation patterns: counts and intervals within conversations for service optimizations. (3) Model response lengths: auto-regressive serving processes of GPT models, showing statistical relations between requests and their responses. (4) System response failures: failures of conversation and API services, showing intensive resource needs and limited availability of LLM services in Azure. The details of the characteristics can serve multiple purposes in LLM serving optimizations, such as system evaluation and trace provisioning. In our demo evaluation with BurstGPT, frequent variations in BurstGPT reveal declines in efficiency, stability, or reliability in realistic LLM serving. We identify that the generalization of KV cache management, scheduling and disaggregation optimizations can be improved under realistic workload evaluations. BurstGPT is publicly available now at https://github.com/HPMLL/BurstGPT and is widely used to develop prototypes of LLM serving frameworks in the industry.

cs.DC↗

arXiv2025

You Only Look One Step: Accelerating Backpropagation in Diffusion Sampling with Gradient Shortcuts

Diffusion models (DMs) have recently demonstrated remarkable success in modeling large-scale data distributions. However, many downstream tasks require guiding the generated content based on specific differentiable metrics, typically necessitating backpropagation during the generation process. This approach is computationally expensive, as generating with DMs often demands tens to hundreds of recursive network calls, resulting in high memory usage and significant time consumption. In this paper, we propose a more efficient alternative that approaches the problem from the perspective of parallel denoising. We show that full backpropagation throughout the entire generation process is unnecessary. The downstream metrics can be optimized by retaining the computational graph of only one step during generation, thus providing a shortcut for gradient propagation. The resulting method, which we call Shortcut Diffusion Optimization (SDO), is generic, high-performance, and computationally lightweight, capable of optimizing all parameter types in diffusion sampling. We demonstrate the effectiveness of SDO on several real-world tasks, including controlling generation by optimizing latent and aligning the DMs by fine-tuning network parameters. Compared to full backpropagation, our approach reduces computational costs by $\sim 90\%$ while maintaining superior performance. Code is available at https://github.com/deng-ai-lab/SDO.

cs.LG↗

arXiv2025

Bridging the Dimensional Chasm: Uncover Layer-wise Dimensional Reduction in Transformers through Token Correlation

The geometric evolution of token representations in large language models (LLMs) presents a fundamental paradox: while human language inherently organizes semantic information in low-dimensional spaces ($\sim 10^1$ dimensions), modern LLMs employ high-dimensional embeddings ($\sim 10^3$ dimensions) processed through Transformer architectures. To resolve this paradox, this work bridges this conceptual gap by developing a geometric framework that tracks token dynamics across Transformers layers. Through layer-wise analysis of intrinsic dimensions across multiple architectures, we reveal an expansion-contraction pattern where tokens diffuse to a "working space" and then progressively project onto lower-dimensional submanifolds. Our finding implies a negative correlation between the working space dimension and parameter-sensitive performance of the LLMs, and indicates that effective models tend to compress tokens into approximately 10-dimensional submanifolds, closely resembling human semantic spaces. This work not only advances LLM interpretability by reframing Transformers layers as projectors that mediate between high-dimensional computation and low-dimensional semantics, but also provides practical tools for model diagnostics that do not rely on task-specific evaluations.

cs.CL↗

arXiv2025

Fourier shape parametrization in covariant density functional theory for nuclear fission

We implement the Fourier shape parametrization within the point-coupling covariant density functional theory to construct the collective space, potential energy surface (PES), and mass tensor, which serve as inputs for the time-dependent generator coordinate method to simulate the fission dynamics. Taking $^{226}$Th as a benchmark, we demonstrate the superiority of Fourier shape parametrization over conventional spherical harmonic parametrization: it significantly enhances the convergence of higher-order collective shape parameters by efficiently characterizing extreme nuclear deformations. Consequently, the new framework generates more reasonable elongated configurations, particularly for the scission configurations, and significantly improves the description of charge distribution near the symmetric fission peak. Moreover, the Fourier shape parametrization provides a smooth and well-defined three-dimensional (3D) PES with minimal correlations between degrees of freedom, enabling high-precision 3D dynamical simulations of fission.

nucl-th↗

arXiv2025

Chern Number Tunable Quantum Anomalous Hall Effect in Compensated Antiferromagnets

We propose to realize the quantum anomalous Hall effect (QAHE) in two-dimensional compensated antiferromagnets without net spin magnetization.} We consider antiferromagnetic MnBi$_2$Te$_4$ as a concrete example. \textcolor{blue}{By breaking the parity-time ($\mathcal{PT}$) symmetry of even-layer MnBi$_2$Te$_4$, we find that the system can host the QAHE with a nonzero Chern number.} We show that by controlling the antiferromagnetic spin configuration, for example, down/up/up/down that breaks $\mathcal{PT}$ symmetry, tetralayer MnBi$_2$Te$_4$ can host a Chern number $\mathcal{C}=-1$. Such spin configuration can be stabilized by pinning the spin orientations of the surfaces. \textcolor{blue}{Furthermore, via tuning the on-site orbital energy and vertical electric fields, we find rich QAHE phases with tunable Chern number of $|\mathcal{C}|=1, 2, 3$. In addition, we reveal that the edge states are layer-selective and primarily locate at the boundaries of the bottom and top layers. Our work not only proposes a scheme to realize Chern number tunable QAHE in antiferromagnets without net spin magnetization, but also provides a platform for layer-selective dissipationless transport devices.

cond-mat.mes-hall↗

arXiv2025

The Layer Hall Effect without External Electric Field

The layer Hall effect is an intriguing phenomenon observed in magnetic topological layered materials, where the Hall response arises from the opposite deflection of electrons on top and bottom layers. To realize layer Hall effect, space-time $\mathcal{PT}$ symmetry is typically broken by applying an external electric field. In this work, we propose a new mechanism to realize the layer Hall effect by introducing inequivalent exchange fields on both surfaces of a topological insulator thin film, in the absence of an electric field. This approach yields a distinct Hall response compared to the conventional electric-field-induced layer Hall effect, particularly with respect to the Fermi level. Taking the topological insulator Sb$_2$Te$_3$ as a concrete example, we demonstrate the feasibility of inducing the layer Hall effect only by coupling the top and bottom surfaces of Sb$_2$Te$_3$ with different magnetic insulators. Notably, we show that both built-in electric-field-induced and inequivalent exchange-fields-induced layer Hall effects can be achieved by tuning the stacking order between Sb$_2$Te$_3$ and the magnetic layers. Given the well-established experimental techniques for fabricating topological insulator thin films, our work offers a viable pathway for realizing layer Hall effect without external electric field.

cond-mat.mes-hall↗

arXiv2025

AI-generated Image Detection: Passive or Watermark?

While text-to-image models offer numerous benefits, they also pose significant societal risks. Detecting AI-generated images is crucial for mitigating these risks. Detection methods can be broadly categorized into passive and watermark-based approaches: passive detectors rely on artifacts present in AI-generated images, whereas watermark-based detectors proactively embed watermarks into such images. A key question is which type of detector performs better in terms of effectiveness, robustness, and efficiency. However, the current literature lacks a comprehensive understanding of this issue. In this work, we aim to bridge that gap by developing ImageDetectBench, the first comprehensive benchmark to compare the effectiveness, robustness, and efficiency of passive and watermark-based detectors. Our benchmark includes four datasets, each containing a mix of AI-generated and non-AI-generated images. We evaluate five passive detectors and four watermark-based detectors against eight types of common perturbations and three types of adversarial perturbations. Our benchmark results reveal several interesting findings. For instance, watermark-based detectors consistently outperform passive detectors, both in the presence and absence of perturbations. Based on these insights, we provide recommendations for detecting AI-generated images, e.g., when both types of detectors are applicable, watermark-based detectors should be the preferred choice. Our code and data are publicly available at https://github.com/moyangkuo/ImageDetectBench.git.

cs.CR↗

arXiv2024

An ab-initio study on engineering quantum anomalous Hall effect in compensated antiferromagnet MnBi$_{2}$Te$_{4}$

Recently, the quantum anomalous Hall effect (QAHE) has been theoretically proposed in compensated antiferromagnetic systems by using the magnetic topological insulator model [see arXiv:2404.13305 (2024)]. However, the related and systematic study based on a realistic material system is still limited. As the only experimentally realized antiferromagnetic topological insulator, MnBi$_{2}$Te$_{4}$ becomes a vital platform for exploring various topological states. In this work, by using the comprehensive first-principles calculations, we demonstrate that the QAHE can also be realized in compensated antiferromagnetic even-septuple-layer MnBi$_{2}$Te$_{4}$ without combined parity-time ($\mathcal{PT}$) symmetry. Using a magnetic topological insulator model, the layer-resolved Chern number is calculated to further understand the physical origin of different Chern numbers. The application of external hydrostatic pressure can strengthen the Te-Te quasicovalent bond due to the dramatic compression of the van der Waals gap. Thus, the resulting topological nontrivial gap can exceed the room-temperature energy scale in a wide range of pressures. Additionally, we find that constructing MnBi$_{2}$Te$_{4}$/CrI$_{3}$ heterostructure can realize the compensated antiferromagnetic configurations with QAHE. Our findings illustrate the realization of QAHE in compensated antiferromagnetic even-septuple-layer MnBi$_{2}$Te$_{4}$ and provide a reliable strategy to obtain the corresponding magnetic configurations.

cond-mat.mtrl-sci↗

arXiv2024

Altermagnetism Induced Topological Phase Transitions in Kane-Mele Model

We theoretically demonstrate that Chern number tunable quantum anomalous Hall effect (QAHE) and second-order topological insulators can be induced in the two-dimensional $\mathbb{Z}_2$ topological insulator (TI), i.e., Kane-Mele model, by applying $d$-wave altermagnetism. When the Néel vector of altermagentism lies in the $x-y$ plane, the $\mathbb{Z}_2$ TI is broken and driven into a second-order topological insulator phase, exhibiting the representative corner states at nanoflakes. When the intrinsic Rashba spin-orbit coupling is further included, the second-order TI is further driven into the QAHE phase with various Chern numbers (e.g., $\mathcal{C}=\pm1$ or $\pm3$). When the Néel vector is along $z$ direction, the intrinsic Rashba spin-orbit coupling is necessary to break the mirror symmetry to allow a sequential emergence of second-order TI and QAHE along with the increase of altermagentism strength. We also observe the QAHE with mixed-chirality, i.e., there exist counter-propagating edge modes but net chiral current at the ribbon boundary. Our work shows that altermagnetism can play a crucial role in exploring a rich variety of topological phases, just like its counterparts of ferromagnetism and antiferromagnetism.

cond-mat.mes-hall↗

arXiv2024

Representational Transfer Learning for Matrix Completion

We propose to transfer representational knowledge from multiple sources to a target noisy matrix completion task by aggregating singular subspaces information. Under our representational similarity framework, we first integrate linear representation information by solving a two-way principal component analysis problem based on a properly debiased matrix-valued dataset. After acquiring better column and row representation estimators from the sources, the original high-dimensional target matrix completion problem is then transformed into a low-dimensional linear regression, of which the statistical efficiency is guaranteed. A variety of extensional arguments, including post-transfer statistical inference and robustness against negative transfer, are also discussed alongside. Finally, extensive simulation results and a number of real data cases are reported to support our claims.

stat.ML↗

arXiv2024

Simultaneous Estimation and Dataset Selection for Transfer Learning in High Dimensions by a Non-convex Penalty

In this paper, we propose to estimate model parameters and identify informative source datasets simultaneously for high-dimensional transfer learning problems with the aid of a non-convex penalty, in contrast to the separate useful dataset selection and transfer learning procedures in the existing literature. To numerically solve the non-convex problem with respect to two specific statistical models, namely the sparse linear regression and the generalized low-rank trace regression models, we adopt the difference of convex (DC) programming with the alternating direction method of multipliers (ADMM) procedures. We theoretically justify the proposed algorithm from both statistical and computational perspectives. Extensive numerical results are reported alongside to validate the theoretical assertions. An \texttt{R} package \texttt{MHDTL} is developed to implement the proposed methods.

stat.ME↗

arXiv2024

Should We Really Edit Language Models? On the Evaluation of Edited Language Models

Model editing has become an increasingly popular alternative for efficiently updating knowledge within language models. Current methods mainly focus on reliability, generalization, and locality, with many methods excelling across these criteria. Some recent works disclose the pitfalls of these editing methods such as knowledge distortion or conflict. However, the general abilities of post-edited language models remain unexplored. In this paper, we perform a comprehensive evaluation on various editing methods and different language models, and have following findings. (1) Existing editing methods lead to inevitable performance deterioration on general benchmarks, indicating that existing editing methods maintain the general abilities of the model within only a few dozen edits. When the number of edits is slightly large, the intrinsic knowledge structure of the model is disrupted or even completely damaged. (2) Instruction-tuned models are more robust to editing, showing less performance drop on general knowledge after editing. (3) Language model with large scale is more resistant to editing compared to small model. (4) The safety of the edited model, is significantly weakened, even for those safety-aligned models. Our findings indicate that current editing methods are only suitable for small-scale knowledge updates within language models, which motivates further research on more practical and reliable editing methods. The details of code and reproduction can be found in https://github.com/lqinfdim/EditingEvaluation.

cs.AI↗

arXiv2024

Applying Color-Kinematics Duality in Pure Yang-Mills at Three Loops

We present the first application of color-kinematics (CK) duality at the three-loop level in non-supersymmetric pure Yang-Mills (YM) theory. Building on the minimal deformation approach introduced in \cite{Li:2023akg}, we extend its use to the three-loop Sudakov form factor. Although three classes of unitarity cuts fail under the globally off-shell CK-dual ansatz, a compact and elegant solution is achieved by deforming a single master numerator. The final numerators exhibit Lorentz invariance in $d$ dimensions and take a local form. This method harnesses CK duality's full potential by enforcing a subset of off-shell dual Jacobi identities for the deformation, offering a promising path toward constructing three-loop amplitudes in non-supersymmetric YM theory and gravity through CK duality and double copy.

hep-th↗

arXiv2024

A data-driven sparse learning approach to reduce chemical reaction mechanisms

Reduction of detailed chemical reaction mechanisms is one of the key methods for mitigating the computational cost of reactive flow simulations. Exploitation of species and elementary reaction sparsity ensures the compactness of the reduced mechanisms. In this work, we propose a novel sparse statistical learning approach for chemical reaction mechanism reduction. Specifically, the reduced mechanism is learned to explicitly reproduce the dynamical evolution of detailed chemical kinetics, while constraining on the sparsity of the reduced reactions at the same time. Compact reduced mechanisms are be achieved as the collection of species that participate in the identified important reactions. We validate our approach by reducing oxidation mechanisms for $n$-heptane (194 species) and 1,3-butadiene (581 species). The results demonstrate that the reduced mechanisms show accurate predictions for the ignition delay times, laminar flame speeds, species mole fraction profiles and turbulence-chemistry interactions across a wide range of operating conditions. Comparative analysis with directed relation graph (DRG)-based methods and the state-of-the-art (SOTA) methods reveals that our sparse learning approach produces reduced mechanisms with fewer species while maintaining the same error limits. The advantages are particularly evident for detailed mechanisms with a larger number of species and reactions. The sparse learning strategy shows significant potential in achieving more substantial reductions in complex chemical reaction mechanisms.

physics.chem-ph↗