Searcharxiv⌕ Search

arXiv subjects

Hao Zhang

Publications and source records attributed to Hao Zhang.

At least 325 records · Page 18Linked to original sources
arXiv2025

Bidirectional Sparse Attention for Faster Video Diffusion Training

Video diffusion Transformer (DiT) models excel in generative quality but hit major computational bottlenecks when producing high-resolution, long-duration videos. The quadratic complexity of full attention leads to prohibitively high training and inference costs. Full attention inefficiency stems from two key challenges: excessive computation due to the inherent sparsity of Queries and Key-Value pairs, and redundant computation as fixed sparse patterns fail to leverage DiT's dynamic attention. To overcome this limitation, we propose a Bidirectional Sparse Attention (BSA) framework for faster video DiT training, the first to dynamically sparsify both Queries and Key-Value pairs within 3D full attention, thereby substantially improving training and inference efficiency. BSA addresses these issues through two key components. Query sparsity is optimized by selecting the most informative query tokens via semantic similarity and with a dynamic spatial-time training strategy, while KV sparsity is achieved by computing a statistical dynamic threshold to retain only the most salient KV blocks for computation. Extensive experiments demonstrate that BSA significantly accelerates DiT training across long sequences, reducing FLOPs by up to 20x and achieving 17.79x faster attention training, while preserving or even surpassing the generative quality of full attention.

cs.CV↗
arXiv2025

Complexity and convergence analysis of a single-loop SDCAM for Lipschitz composite optimization and beyond

We develop and analyze a single-loop algorithm for minimizing the sum of a Lipschitz differentiable function $f$, a prox-friendly proper closed function $g$ (with a closed domain on which $g$ is continuous) and the composition of another prox-friendly proper closed function $h$ (whose domain is closed on which $h$ is continuous) with a continuously differentiable mapping $c$ (that is Lipschitz continuous and Lipschitz differentiable on the convex closure of the domain of $g$). Such models arise naturally in many contemporary applications, where $f$ is the loss function for data misfit, and $g$ and $h$ are nonsmooth functions for inducing desirable structures in $x$ and $c(x)$. Existing single-loop algorithms mainly focus either on the case where $h$ is Lipschitz continuous or the case where $h$ is an indicator function of a closed convex set. In this paper, we develop a single-loop algorithm for more general possibly non-Lipschitz $h$. Our algorithm is a single-loop variant of the successive difference-of-convex approximation method (SDCAM) proposed in [22]. We show that when $h$ is Lipschitz, our algorithm exhibits an iteration complexity that matches the best known complexity result for obtaining an $(ε_1,ε_2,0)$-stationary point. Moreover, we show that, by assuming additionally that dom $g$ is compact, our algorithm exhibits an iteration complexity of $\tilde{O}(ε^{-4})$ for obtaining an $(ε,ε,ε)$-stationary point when $h$ is merely continuous and real-valued. Furthermore, we consider a scenario where $h$ does not have full domain and establish vanishing bounds on successive changes of iterates. Finally, in all three cases mentioned above, we show that one can construct a subsequence such that any accumulation point $x^*$ satisfies $c(x^*)\in$ dom $h$, and if a standard constraint qualification holds at $x^*$, then $x^*$ is a stationary point.

math.OC↗
arXiv2025

HY-Motion 1.0: Scaling Flow Matching Models for Text-To-Motion Generation

We present HY-Motion 1.0, a series of state-of-the-art, large-scale, motion generation models capable of generating 3D human motions from textual descriptions. HY-Motion 1.0 represents the first successful attempt to scale up Diffusion Transformer (DiT)-based flow matching models to the billion-parameter scale within the motion generation domain, delivering instruction-following capabilities that significantly outperform current open-source benchmarks. Uniquely, we introduce a comprehensive, full-stage training paradigm -- including large-scale pretraining on over 3,000 hours of motion data, high-quality fine-tuning on 400 hours of curated data, and reinforcement learning from both human feedback and reward models -- to ensure precise alignment with the text instruction and high motion quality. This framework is supported by our meticulous data processing pipeline, which performs rigorous motion cleaning and captioning. Consequently, our model achieves the most extensive coverage, spanning over 200 motion categories across 6 major classes. We release HY-Motion 1.0 to the open-source community to foster future research and accelerate the transition of 3D human motion generation models towards commercial maturity.

cs.CV↗
arXiv2025

Re-Visible Dual-Domain Self-Supervised Deep Unfolding Network for MRI Reconstruction

Magnetic Resonance Imaging (MRI) is widely used in clinical practice, but suffered from prolonged acquisition time. Although deep learning methods have been proposed to accelerate acquisition and demonstrate promising performance, they rely on high-quality fully-sampled datasets for training in a supervised manner. However, such datasets are time-consuming and expensive-to-collect, which constrains their broader applications. On the other hand, self-supervised methods offer an alternative by enabling learning from under-sampled data alone, but most existing methods rely on further partitioned under-sampled k-space data as model's input for training, resulting in a loss of valuable information. Additionally, their models have not fully incorporated image priors, leading to degraded reconstruction performance. In this paper, we propose a novel re-visible dual-domain self-supervised deep unfolding network to address these issues when only under-sampled datasets are available. Specifically, by incorporating re-visible dual-domain loss, all under-sampled k-space data are utilized during training to mitigate information loss caused by further partitioning. This design enables the model to implicitly adapt to all under-sampled k-space data as input. Additionally, we design a deep unfolding network based on Chambolle and Pock Proximal Point Algorithm (DUN-CP-PPA) to achieve end-to-end reconstruction, incorporating imaging physics and image priors to guide the reconstruction process. By employing a Spatial-Frequency Feature Extraction (SFFE) block to capture global and local feature representation, we enhance the model's efficiency to learn comprehensive image priors. Experiments conducted on the fastMRI and IXI datasets demonstrate that our method significantly outperforms state-of-the-art approaches in terms of reconstruction performance.

eess.IV↗
arXiv2025

PTalker: Personalized Speech-Driven 3D Talking Head Animation via Style Disentanglement and Modality Alignment

Speech-driven 3D talking head generation aims to produce lifelike facial animations precisely synchronized with speech. While considerable progress has been made in achieving high lip-synchronization accuracy, existing methods largely overlook the intricate nuances of individual speaking styles, which limits personalization and realism. In this work, we present a novel framework for personalized 3D talking head animation, namely "PTalker". This framework preserves speaking style through style disentanglement from audio and facial motion sequences and enhances lip-synchronization accuracy through a three-level alignment mechanism between audio and mesh modalities. Specifically, to effectively disentangle style and content, we design disentanglement constraints that encode driven audio and motion sequences into distinct style and content spaces to enhance speaking style representation. To improve lip-synchronization accuracy, we adopt a modality alignment mechanism incorporating three aspects: spatial alignment using Graph Attention Networks to capture vertex connectivity in the 3D mesh structure, temporal alignment using cross-attention to capture and synchronize temporal dependencies, and feature alignment by top-k bidirectional contrastive losses and KL divergence constraints to ensure consistency between speech and mesh modalities. Extensive qualitative and quantitative experiments on public datasets demonstrate that PTalker effectively generates realistic, stylized 3D talking heads that accurately match identity-specific speaking styles, outperforming state-of-the-art methods. The source code and supplementary videos are available at: PTalker.

cs.CV↗
arXiv2025

Self-motion as a structural prior for coherent and robust formation of cognitive maps

Most computational accounts of cognitive maps assume that stability is achieved primarily through sensory anchoring, with self-motion contributing to incremental positional updates only. However, biological spatial representations often remain coherent even when sensory cues degrade or conflict, suggesting that self-motion may play a deeper organizational role. Here, we show that self-motion can act as a structural prior that actively organizes the geometry of learned cognitive maps. We embed a path-integration-based motion prior in a predictive-coding framework, implemented using a capacity-efficient, brain-inspired recurrent mechanism combining spiking dynamics, analog modulation and adaptive thresholds. Across highly aliased, dynamically changing and naturalistic environments, this structural prior consistently stabilizes map formation, improving local topological fidelity, global positional accuracy and next-step prediction under sensory ambiguity. Mechanistic analyses reveal that the motion prior itself encodes geometrically precise trajectories under tight constraints of internal states and generalizes zero-shot to unseen environments, outperforming simpler motion-based constraints. Finally, deployment on a quadrupedal robot demonstrates that motion-derived structural priors enhance online landmark-based navigation under real-world sensory variability. Together, these results reframe self-motion as an organizing scaffold for coherent spatial representations, showing how brain-inspired principles can systematically strengthen spatial intelligence in embodied artificial agents.

q-bio.NC↗
arXiv2025

AsyMoE: Leveraging Modal Asymmetry for Enhanced Expert Specialization in Large Vision-Language Models

Large Vision-Language Models (LVLMs) have demonstrated impressive performance on multimodal tasks through scaled architectures and extensive training. However, existing Mixture of Experts (MoE) approaches face challenges due to the asymmetry between visual and linguistic processing. Visual information is spatially complete, while language requires maintaining sequential context. As a result, MoE models struggle to balance modality-specific features and cross-modal interactions. Through systematic analysis, we observe that language experts in deeper layers progressively lose contextual grounding and rely more on parametric knowledge rather than utilizing the provided visual and linguistic information. To address this, we propose AsyMoE, a novel architecture that models this asymmetry using three specialized expert groups. We design intra-modality experts for modality-specific processing, hyperbolic inter-modality experts for hierarchical cross-modal interactions, and evidence-priority language experts to suppress parametric biases and maintain contextual grounding. Extensive experiments demonstrate that AsyMoE achieves 26.58% and 15.45% accuracy improvements over vanilla MoE and modality-specific MoE respectively, with 25.45% fewer activated parameters than dense models.

cs.CV↗
arXiv2025

GraphGeo: Multi-Agent Debate Framework for Visual Geo-localization with Heterogeneous Graph Neural Networks

Visual geo-localization requires extensive geographic knowledge and sophisticated reasoning to determine image locations without GPS metadata. Traditional retrieval methods are constrained by database coverage and quality. Recent Large Vision-Language Models (LVLMs) enable direct location reasoning from image content, yet individual models struggle with diverse geographic regions and complex scenes. Existing multi-agent systems improve performance through model collaboration but treat all agent interactions uniformly. They lack mechanisms to handle conflicting predictions effectively. We propose \textbf{GraphGeo}, a multi-agent debate framework using heterogeneous graph neural networks for visual geo-localization. Our approach models diverse debate relationships through typed edges, distinguishing supportive collaboration, competitive argumentation, and knowledge transfer. We introduce a dual-level debate mechanism combining node-level refinement and edge-level argumentation modeling. A cross-level topology refinement strategy enables co-evolution between graph structure and agent representations. Experiments on multiple benchmarks demonstrate GraphGeo significantly outperforms state-of-the-art methods. Our framework transforms cognitive conflicts between agents into enhanced geo-localization accuracy through structured debate.

cs.CV↗
arXiv2025

G2rammar: Bilingual Grammar Modeling for Enhanced Text-attributed Graph Learning

Text-attributed graphs require models to effectively integrate both structural topology and semantic content. Recent approaches apply large language models to graphs by linearizing structures into token sequences through random walks. These methods create concise graph vocabularies to replace verbose natural language descriptions. However, they overlook a critical component that makes language expressive: grammar. In natural language, grammar assigns syntactic roles to words and defines their functions within sentences. Similarly, nodes in graphs play distinct structural roles as hubs, bridges, or peripheral members. Current graph language methods provide tokens without grammatical annotations to indicate these structural or semantic roles. This absence limits language models' ability to reason about graph topology effectively. We propose \textbf{G2rammar}, a bilingual grammar framework that explicitly encodes both structural and semantic grammar for text-attributed graphs. Structural grammar characterizes topological roles through centrality and neighborhood patterns. Semantic grammar captures content relationships through textual informativity. The framework implements two-stage learning with structural grammar pre-training followed by semantic grammar fine-tuning. Extensive experiments on real-world datasets demonstrate that G2rammar consistently outperforms competitive baselines by providing language models with the grammatical context needed to understand graph structures.

cs.GR↗
arXiv2025

CETCAM: Camera-Controllable Video Generation via Consistent and Extensible Tokenization

Achieving precise camera control in video generation remains challenging, as existing methods often rely on camera pose annotations that are difficult to scale to large and dynamic datasets and are frequently inconsistent with depth estimation, leading to train-test discrepancies. We introduce CETCAM, a camera-controllable video generation framework that eliminates the need for camera annotations through a consistent and extensible tokenization scheme. CETCAM leverages recent advances in geometry foundation models, such as VGGT, to estimate depth and camera parameters and converts them into unified, geometry-aware tokens. These tokens are seamlessly integrated into a pretrained video diffusion backbone via lightweight context blocks. Trained in two progressive stages, CETCAM first learns robust camera controllability from diverse raw video data and then refines fine-grained visual quality using curated high-fidelity datasets. Extensive experiments across multiple benchmarks demonstrate state-of-the-art geometric consistency, temporal stability, and visual realism. Moreover, CETCAM exhibits strong adaptability to additional control modalities, including inpainting and layout control, highlighting its flexibility beyond camera control. The project page is available at https://sjtuytc.github.io/CETCam_project_page.github.io/.

cs.CV↗
arXiv2025

Reactive near-field subwavelength microwave imaging with a non-invasive Rydberg probe

Non-invasive microwave field imaging--accurately mapping field distributions without perturbing them--is essential in areas such as aerospace engineering, biomedical imaging and integrated-circuit diagnostics. Conventional metal probes, however, inevitably perturb reactive near fields: they act as strong scatterers that drive induced currents and secondary radiation, remap evanescent components and thereby degrade both accuracy and spatial resolution, particularly in the reactive near-field regime that is most relevant to these applications. Here we demonstrate, to our knowledge for the first time, reactive near-field subwavelength imaging of microwave fields using the quantum non-demolition properties of Rydberg atoms, realized with a compact, non-invasive single-ended fibre-integrated Rydberg probe engineered to minimize field disturbance. The probe achieves an imaging resolution of {\unboldmath$λ/56$}, and the measured field distributions agree with full-wave simulations with structural similarity approaching unity, confirming both its subwavelength spatial resolution and its genuinely non-invasive character compared with conventional metal-based probes. Because the atomic sensor is intrinsically isotropic, the same device can faithfully image multi-dimensional field structures without orientation-dependent calibration. Our results therefore establish a general, non-invasive route to high-accuracy, subwavelength reactive near-field microwave imaging, with particular promise for applications such as chip-defect detection and integrated-circuit diagnostics, where even small perturbations by the probe can mask the underlying physics of interest.

quant-ph↗
arXiv2025

Fast and Accurate Causal Parallel Decoding using Jacobi Forcing

Multi-token generation has emerged as a promising paradigm for accelerating transformer-based large model inference. Recent efforts primarily explore diffusion Large Language Models (dLLMs) for parallel decoding to reduce inference latency. To achieve AR-level generation quality, many techniques adapt AR models into dLLMs to enable parallel decoding. However, they suffer from limited speedup compared to AR models due to a pretrain-to-posttrain mismatch. Specifically, the masked data distribution in post-training deviates significantly from the real-world data distribution seen during pretraining, and dLLMs rely on bidirectional attention, which conflicts with the causal prior learned during pretraining and hinders the integration of exact KV cache reuse. To address this, we introduce Jacobi Forcing, a progressive distillation paradigm where models are trained on their own generated parallel decoding trajectories, smoothly shifting AR models into efficient parallel decoders while preserving their pretrained causal inference property. The models trained under this paradigm, Jacobi Forcing Model, achieves 3.8x wall-clock speedup on coding and math benchmarks with minimal loss in performance. Based on Jacobi Forcing Models' trajectory characteristics, we introduce multi-block decoding with rejection recycling, which enables up to 4.5x higher token acceptance count per iteration and nearly 4.0x wall-clock speedup, effectively trading additional compute for lower inference latency. Our code is available at https://github.com/hao-ai-lab/JacobiForcing.

cs.CL↗
arXiv2025

SpectrumFM: A Foundation Model for Intelligent Spectrum Management

Intelligent spectrum management is crucial for improving spectrum efficiency and achieving secure utilization of spectrum resources. However, existing intelligent spectrum management methods, typically based on small-scale models, suffer from notable limitations in recognition accuracy, convergence speed, and generalization, particularly in the complex and dynamic spectrum environments. To address these challenges, this paper proposes a novel spectrum foundation model, termed SpectrumFM, establishing a new paradigm for spectrum management. SpectrumFM features an innovative encoder architecture that synergistically exploits the convolutional neural networks and the multi-head self-attention mechanisms to enhance feature extraction and enable robust representation learning. The model is pre-trained via two novel self-supervised learning tasks, namely masked reconstruction and next-slot signal prediction, which leverage large-scale in-phase and quadrature (IQ) data to achieve comprehensive and transferable spectrum representations. Furthermore, a parameter-efficient fine-tuning strategy is proposed to enable SpectrumFM to adapt to various downstream spectrum management tasks, including automatic modulation classification (AMC), wireless technology classification (WTC), spectrum sensing (SS), and anomaly detection (AD). Extensive experiments demonstrate that SpectrumFM achieves superior performance in terms of accuracy, robustness, adaptability, few-shot learning efficiency, and convergence speed, consistently outperforming conventional methods across multiple benchmarks. Specifically, SpectrumFM improves AMC accuracy by up to 12.1% and WTC accuracy by 9.3%, achieves an area under the curve (AUC) of 0.97 in SS at -4 dB signal-to-noise ratio (SNR), and enhances AD performance by over 10%.

eess.SP↗
arXiv2025

Theory of Remaining Exceptional Points from Nongeneric Splitting in Non-Hermitian Systems

In non-Hermitian physics, high-order exceptional points(HOEPs) with eigenvalues and eigenvectors coalesce are known for their enhanced sensitivity to perturbations. Typically, they exhibit eigenvalue splitting that scales as ε^(1/n), which is referred to as the generic response. However, under certain conditions, a nongeneric response of HOEPs occurs where the splitting follows a lower order ε^(1/m) (m<n). A nongeneric response of HOEPs with a lower order splitting lead to the remaining EPs. While the presence of these remaining EPs is acknowledged, a thorough elucidation of their fundamental properties has yet to be achieved. In this work, we demonstrate those unsplit eigenvalue points must constitute remaining EPs in a perturbed n-orders HOEPs system. Combining graph theory and topological analysis, the number and splitting order of the remaining EPs is studied. This framework not only resolves a fundamental challenge in HOEPs but also paves the way for exploiting remaining EPs in applications such as anisotropic sensing and the design of Dirac exceptional points.

physics.optics↗
arXiv2025

PDTrim: Targeted Pruning for Prefill-Decode Disaggregation in Inference

Large Language Models (LLMs) demonstrate exceptional capabilities across various tasks, but their deployment is constrained by high computational and memory costs. Model pruning provides an effective means to alleviate these demands. However, existing methods often ignore the characteristics of prefill-decode (PD) disaggregation in practice. In this paper, we propose a pruning method that is highly integrated with PD disaggregation, enabling more precise pruning of blocks. Our approach constructs pruning and distillation sets to perform iterative block removal, obtaining better pruning solutions. Moreover, we analyze the pruning sensitivity of the prefill and decode stages and identify removable blocks specific to each stage, making it well suited for PD disaggregation deployment. Extensive experiments demonstrate our approach consistently achieves strong performance in both PD disaggregation and PD unified (non-PD disaggregation) settings, and can also be extended to other non-block pruning methods. Under the same settings, our method achieves improved performance and faster inference.

cs.CL↗
arXiv2025

HyperAdaLoRA: Accelerating LoRA Rank Allocation During Training via Hypernetworks without Sacrificing Performance

Parameter-Efficient Fine-Tuning (PEFT), especially Low-Rank Adaptation (LoRA), has emerged as a promising approach to fine-tuning large language models(LLMs) while reducing computational and memory overhead. However, LoRA assumes a uniform rank \textit{r} for each incremental matrix, not accounting for the varying significance of weight matrices across different modules and layers. AdaLoRA leverages Singular Value Decomposition (SVD) to parameterize updates and employs pruning of singular values to introduce dynamic rank allocation, thereby enhancing adaptability. However, during the training process, it often encounters issues of slow convergence speed and high computational overhead. To address this issue, we propose HyperAdaLoRA, a novel framework that accelerates the convergence of AdaLoRA by leveraging a hypernetwork. Instead of directly optimizing the components of Singular Value Decomposition $(P, Λ, Q)$, HyperAdaLoRA employs a hypernetwork based on attention mechanisms to dynamically generate these parameters. By pruning the outputs of the hypernetwork that generates the singular values, dynamic rank allocation is achieved. Comprehensive experiments on various datasets and models demonstrate that our method achieves faster convergence without sacrificing performance. Additionally, further extension experiments on other LoRA-based approaches validate the broad applicability of our method.

cs.LG↗
arXiv2025

Zero-shot 3D Map Generation with LLM Agents: A Dual-Agent Architecture for Procedural Content Generation

Procedural Content Generation (PCG) offers scalable methods for algorithmically creating complex, customizable worlds. However, controlling these pipelines requires the precise configuration of opaque technical parameters. We propose a training-free architecture that utilizes LLM agents for zero-shot PCG parameter configuration. While Large Language Models (LLMs) promise a natural language interface for PCG tools, off-the-shelf models often fail to bridge the semantic gap between abstract user instructions and strict parameter specifications. Our system pairs an Actor agent with a Critic agent, enabling an iterative workflow where the system autonomously reasons over tool parameters and refines configurations to progressively align with human design preferences. We validate this approach on the generation of various 3D maps, establishing a new benchmark for instruction-following in PCG. Experiments demonstrate that our approach outperforms single-agent baselines, producing diverse and structurally valid environments from natural language descriptions. These results demonstrate that off-the-shelf LLMs can be effectively repurposed as generalized agents for arbitrary PCG tools. By shifting the burden from model training to architectural reasoning, our method offers a scalable framework for mastering complex software without task-specific fine-tuning.

cs.AI↗