Searcharxiv⌕ Search

arXiv subjects

Hao Zhang

Publications and source records attributed to Hao Zhang.

At least 397 records · Page 22Linked to original sources
arXiv2025

Fast-dLLM v2: Efficient Block-Diffusion LLM

Autoregressive (AR) large language models (LLMs) have achieved remarkable performance across a wide range of natural language tasks, yet their inherent sequential decoding limits inference efficiency. In this work, we propose Fast-dLLM v2, a carefully designed block diffusion language model (dLLM) that efficiently adapts pretrained AR models into dLLMs for parallel text generation, requiring only approximately 1B tokens of fine-tuning. This represents a 500x reduction in training data compared to full-attention diffusion LLMs such as Dream (580B tokens), while preserving the original model's performance. Our approach introduces a novel training recipe that combines a block diffusion mechanism with a complementary attention mask, enabling blockwise bidirectional context modeling without sacrificing AR training objectives. To further accelerate decoding, we design a hierarchical caching mechanism: a block-level cache that stores historical context representations across blocks, and a sub-block cache that enables efficient parallel generation within partially decoded blocks. Coupled with our parallel decoding pipeline, Fast-dLLM v2 achieves up to 2.5x speedup over standard AR decoding without compromising generation quality. Extensive experiments across diverse benchmarks demonstrate that Fast-dLLM v2 matches or surpasses AR baselines in accuracy, while delivering state-of-the-art efficiency among dLLMs - marking a significant step toward the practical deployment of fast and accurate LLMs. Code and model will be publicly released.

cs.CL↗
arXiv2025

ASIA: Adaptive 3D Segmentation using Few Image Annotations

We introduce ASIA (Adaptive 3D Segmentation using few Image Annotations), a novel framework that enables segmentation of possibly non-semantic and non-text-describable "parts" in 3D. Our segmentation is controllable through a few user-annotated in-the-wild images, which are easier to collect than multi-view images, less demanding to annotate than 3D models, and more precise than potentially ambiguous text descriptions. Our method leverages the rich priors of text-to-image diffusion models, such as Stable Diffusion (SD), to transfer segmentations from image space to 3D, even when the annotated and target objects differ significantly in geometry or structure. During training, we optimize a text token for each segment and fine-tune our model with a novel cross-view part correspondence loss. At inference, we segment multi-view renderings of the 3D mesh, fuse the labels in UV-space via voting, refine them with our novel Noise Optimization technique, and finally map the UV-labels back onto the mesh. ASIA provides a practical and generalizable solution for both semantic and non-semantic 3D segmentation tasks, outperforming existing methods by a noticeable margin in both quantitative and qualitative evaluations.

cs.CV↗
arXiv2025

Emission-GPT: A domain-specific language model agent for knowledge retrieval, emission inventory and data analysis

Improving air quality and addressing climate change relies on accurate understanding and analysis of air pollutant and greenhouse gas emissions. However, emission-related knowledge is often fragmented and highly specialized, while existing methods for accessing and compiling emissions data remain inefficient. These issues hinder the ability of non-experts to interpret emissions information, posing challenges to research and management. To address this, we present Emission-GPT, a knowledge-enhanced large language model agent tailored for the atmospheric emissions domain. Built on a curated knowledge base of over 10,000 documents (including standards, reports, guidebooks, and peer-reviewed literature), Emission-GPT integrates prompt engineering and question completion to support accurate domain-specific question answering. Emission-GPT also enables users to interactively analyze emissions data via natural language, such as querying and visualizing inventories, analyzing source contributions, and recommending emission factors for user-defined scenarios. A case study in Guangdong Province demonstrates that Emission-GPT can extract key insights--such as point source distributions and sectoral trends--directly from raw data with simple prompts. Its modular and extensible architecture facilitates automation of traditionally manual workflows, positioning Emission-GPT as a foundational tool for next-generation emission inventory development and scenario-based assessment.

cs.CL↗
arXiv2025

In-2-4D: Inbetweening from Two Single-View Images to 4D Generation

We pose a new problem, In-2-4D, for generative 4D (i.e., 3D + motion) inbetweening to interpolate two single-view images. In contrast to video/4D generation from only text or a single image, our interpolative task can leverage more precise motion control to better constrain the generation. Given two monocular RGB images representing the start and end states of an object in motion, our goal is to generate and reconstruct the motion in 4D, without making assumptions on the object category, motion type, length, or complexity. To handle such arbitrary and diverse motions, we utilize a foundational video interpolation model for motion prediction. However, large frame-to-frame motion gaps can lead to ambiguous interpretations. To this end, we employ a hierarchical approach to identify keyframes that are visually close to the input states while exhibiting significant motions, then generate smooth fragments between them. For each fragment, we construct a 3D representation of the keyframe using Gaussian Splatting (3DGS). The temporal frames within the fragment guide the motion, enabling their transformation into dynamic 3DGS through a deformation field. To improve temporal consistency and refine the 3D motion, we expand the self-attention of multi-view diffusion across timesteps and apply rigid transformation regularization. Finally, we merge the independently generated 3D motion segments by interpolating boundary deformation fields and optimizing them to align with the guiding video, ensuring smooth and flicker-free transitions. Through extensive qualitative and quantitive experiments as well as a user study, we demonstrate the effectiveness of our method and design choices.

cs.GR↗
arXiv2025

MagicTryOn: Harnessing Diffusion Transformer for Garment-Preserving Video Virtual Try-on

Video Virtual Try-On (VVT) aims to synthesize garments that appear natural across consecutive video frames, capturing both their dynamics and interactions with human motion. Despite recent progress, existing VVT methods still suffer from inadequate garment fidelity and limited spatiotemporal consistency. The reasons are: (1) under-exploitation of garment information, with limited garment cues being injected, resulting in weaker fine-detail fidelity; and (2) a lack of spatiotemporal modeling, which hampers cross-frame identity consistency and causes temporal jitter and appearance drift. In this paper, we present MagicTryOn, a diffusion-transformer based framework for garment-preserving video virtual try-on. To preserve fine-grained garment details, we propose a fine-grained garment-preservation strategy that disentangles garment cues and injects these decomposed priors into the denoising process. To improve temporal garment consistency and suppress jitter, we introduce a garment-aware spatiotemporal rotary positional embedding (RoPE) that extends RoPE within full self-attention, using spatiotemporal relative positions to modulate garment tokens. We further impose a mask-aware loss during training to enhance fidelity within garment regions. Moreover, we adopt distribution-matching distillation to compress the sampling trajectory to four steps, enabling real-time inference without degrading garment fidelity. Extensive quantitative and qualitative experiments demonstrate that MagicTryOn outperforms existing methods, delivering superior garment-detail fidelity and temporal stability in unconstrained settings.

cs.CV↗
arXiv2025

ART-DECO: Arbitrary Text Guidance for 3D Detailizer Construction

We introduce a 3D detailizer, a neural model which can instantaneously (in <1s) transform a coarse 3D shape proxy into a high-quality asset with detailed geometry and texture as guided by an input text prompt. Our model is trained using the text prompt, which defines the shape class and characterizes the appearance and fine-grained style of the generated details. The coarse 3D proxy, which can be easily varied and adjusted (e.g., via user editing), provides structure control over the final shape. Importantly, our detailizer is not optimized for a single shape; it is the result of distilling a generative model, so that it can be reused, without retraining, to generate any number of shapes, with varied structures, whose local details all share a consistent style and appearance. Our detailizer training utilizes a pretrained multi-view image diffusion model, with text conditioning, to distill the foundational knowledge therein into our detailizer via Score Distillation Sampling (SDS). To improve SDS and enable our detailizer architecture to learn generalizable features over complex structures, we train our model in two training stages to generate shapes with increasing structural complexity. Through extensive experiments, we show that our method generates shapes of superior quality and details compared to existing text-to-3D models under varied structure control. Our detailizer can refine a coarse shape in less than a second, making it possible to interactively author and adjust 3D shapes. Furthermore, the user-imposed structure control can lead to creative, and hence out-of-distribution, 3D asset generations that are beyond the current capabilities of leading text-to-3D generative models. We demonstrate an interactive 3D modeling workflow our method enables, and its strong generalizability over styles, structures, and object categories.

cs.GR↗
arXiv2025

MMR1: Enhancing Multimodal Reasoning with Variance-Aware Sampling and Open Resources

Large multimodal reasoning models have achieved rapid progress, but their advancement is constrained by two major limitations: the absence of open, large-scale, high-quality long chain-of-thought (CoT) data, and the instability of reinforcement learning (RL) algorithms in post-training. Group Relative Policy Optimization (GRPO), the standard framework for RL fine-tuning, is prone to gradient vanishing when reward variance is low, which weakens optimization signals and impairs convergence. This work makes three contributions: (1) We propose Variance-Aware Sampling (VAS), a data selection strategy guided by Variance Promotion Score (VPS) that combines outcome variance and trajectory diversity to promote reward variance and stabilize policy optimization. (2) We release large-scale, carefully curated resources containing ~1.6M long CoT cold-start data and ~15k RL QA pairs, designed to ensure quality, difficulty, and diversity, along with a fully reproducible end-to-end training codebase. (3) We open-source a family of multimodal reasoning models in multiple scales, establishing standardized baselines for the community. Experiments across mathematical reasoning benchmarks demonstrate the effectiveness of both the curated data and the proposed VAS. Comprehensive ablation studies and analyses provide further insight into the contributions of each component. In addition, we theoretically establish that reward variance lower-bounds the expected policy gradient magnitude, with VAS serving as a practical mechanism to realize this guarantee. Our code, data, and checkpoints are available at https://github.com/LengSicong/MMR1.

cs.CV↗
arXiv2025

Two ADI compact difference methods for variable-exponent diffusion wave equations

In this work, we study two-dimensional diffusion-wave equations with variable exponent, modeling mechanical diffusive wave propagation in viscoelastic media with spatially varying properties. We first transform the diffusion-wave model into an equivalent form via the convolution method. Two time discretization strategies are then applied to approximate each term in the transformed equation, yielding two fully discrete schemes based on a spatial compact finite difference method. To reduce computational cost, the alternating direction implicit (ADI) technique is employed. We prove that both ADI compact schemes are unconditionally stable and convergent. Under solution regularity, the first scheme achieves $α(0)$-order accuracy in time and fourth-order accuracy in space, while the second scheme attains second-order accuracy in time and fourth-order accuracy in space. Numerical experiments confirm the theoretical error estimates and demonstrate the efficiency of the proposed methods.

math.NA↗
arXiv2025

GeoPQA: Bridging the Visual Perception Gap in MLLMs for Geometric Reasoning

Recent advancements in reinforcement learning (RL) have enhanced the reasoning abilities of large language models (LLMs), yet the impact on multimodal LLMs (MLLMs) is limited. Particularly in vision-intensive tasks like geometric reasoning, MLLMs hallucinate frequently, leading to inaccurate reasoning. We attribute this to the perceptual bottleneck in MLLMs, which caps the benefits of reasoning training. To quantify this, we design a Geo-Perception Question-Answering (GeoPQA) benchmark, targeting basic geometric concepts and spatial relationships. Experiments on GeoPQA reveal significant shortcomings of MLLMs in visual perception, which constrain RL reward signals for effective training. To address this bottleneck, we propose a two-stage RL training framework by first enhancing the visual perception of geometric structures, then fostering reasoning capabilities. Applied to Qwen2.5-VL-3B-Instruct, our two-stage training improves geometric reasoning by 9.7% and geometric problem solving by 9.1%, compared to the direct reasoning training approach. Our method also generalizes to other vision-intensive domains like figure understanding, highlighting the importance of perceptual grounding in effective MLLM reasoning.

cs.CL↗
arXiv2025

Sugar-Coated Poison: Benign Generation Unlocks LLM Jailbreaking

With the increasingly deep integration of large language models (LLMs) across diverse domains, the effectiveness of their safety mechanisms is encountering severe challenges. Currently, jailbreak attacks based on prompt engineering have become a major safety threat. However, existing methods primarily rely on black-box manipulation of prompt templates, resulting in poor interpretability and limited generalization. To break through the bottleneck, this study first introduces the concept of Defense Threshold Decay (DTD), revealing the potential safety impact caused by LLMs' benign generation: as benign content generation in LLMs increases, the model's focus on input instructions progressively diminishes. Building on this insight, we propose the Sugar-Coated Poison (SCP) attack paradigm, which uses a "semantic reversal" strategy to craft benign inputs that are opposite in meaning to malicious intent. This strategy induces the models to generate extensive benign content, thereby enabling adversarial reasoning to bypass safety mechanisms. Experiments show that SCP outperforms existing baselines. Remarkably, it achieves an average attack success rate of 87.23% across six LLMs. For defense, we propose Part-of-Speech Defense (POSD), leveraging verb-noun dependencies for syntactic analysis to enhance safety of LLMs while preserving their generalization ability.

cs.CR↗
arXiv2025

FSOS-AMC: Few-Shot Open-Set Learning for Automatic Modulation Classification Over Multipath Fading Channels

Automatic modulation classification (AMC) plays a vital role in advancing future wireless communication networks. Although deep learning (DL)-based AMC frameworks have demonstrated remarkable classification capabilities, they typically require large-scale training datasets and assume consistent class distributions between training and testing data-prerequisites that prove challenging in few-shot and open-set scenarios. To address these limitations, we propose a novel few-shot open-set AMC (FSOS-AMC) framework that integrates a multisequence multiscale attention network (MS-MSANet), meta-prototype training, and a modular open-set classifier. The MS-MSANet extracts features from multisequence input signals, while meta-prototype training optimizes both the feature extractor and the modular open-set classifier, which can effectively categorize testing data into known modulation types or identify potential unknown modulations. Extensive simulation results demonstrate that our FSOS-AMC framework achieves superior performance in few-shot open-set scenarios compared to state-of-the-art methods. Specifically, the framework exhibits higher classification accuracy for both known and unknown modulations, as validated by improved accuracy and area under the receiver operating characteristic curve (AUROC) metrics. Moreover, the proposed framework demonstrates remarkable robustness under challenging low signal-to-noise ratio (SNR) conditions, significantly outperforming existing approaches.

eess.SP↗
arXiv2025

RGB-Only Supervised Camera Parameter Optimization in Dynamic Scenes

Although COLMAP has long remained the predominant method for camera parameter optimization in static scenes, it is constrained by its lengthy runtime and reliance on ground truth (GT) motion masks for application to dynamic scenes. Many efforts attempted to improve it by incorporating more priors as supervision such as GT focal length, motion masks, 3D point clouds, camera poses, and metric depth, which, however, are typically unavailable in casually captured RGB videos. In this paper, we propose a novel method for more accurate and efficient camera parameter optimization in dynamic scenes solely supervised by a single RGB video, dubbed ROS-Cam. Our method consists of three key components: (1) Patch-wise Tracking Filters, to establish robust and maximally sparse hinge-like relations across the RGB video. (2) Outlier-aware Joint Optimization, for efficient camera parameter optimization by adaptive down-weighting of moving outliers, without reliance on motion priors. (3) A Two-stage Optimization Strategy, to enhance stability and optimization speed by a trade-off between the Softplus limits and convex minima in losses. We visually and numerically evaluate our camera estimates. To further validate accuracy, we feed the camera estimates into a 4D reconstruction method and assess the resulting 3D scenes, and rendered 2D RGB and depth maps. We perform experiments on 4 real-world datasets (NeRF-DS, DAVIS, iPhone, and TUM-dynamics) and 1 synthetic dataset (MPI-Sintel), demonstrating that our method estimates camera parameters more efficiently and accurately with a single RGB video as the only supervision.

cs.CV↗
arXiv2025

Distributed Multi-Task Learning for Joint Wireless Signal Enhancement and Recognition

Wireless signal recognition (WSR) is crucial in modern and future wireless communication networks since it aims to identify the properties of the received signal in a no-collaborative manner. However, it is challenging to accurately classify signals in low signal-to-noise ratio (SNR) conditions and distributed network settings. In this paper, we propose a novel distributed multi-task learning framework for joint wireless signal enhancement and recognition (WSER), addressing the crucial need for non-collaborative signal identification in modern wireless networks. Our approach integrates a wireless signal enhancement and recognition network (WSERNet) with FedProx+, an enhanced federated learning algorithm designed for heterogeneous data distributions. Specifically, WSERNet leverages an asymmetric convolution block (ACBlock) to capture long-range dependencies in the input signal and improve the performance of the deep learning model. FedProx+ introduces a proximal term to the loss function to encourage the model updates to be closer to the previous model, enhancing the convergence speed and robustness of federated learning. Extensive experiments demonstrate the effectiveness of the proposed framework for joint WSER, achieving superior performance compared to state-of-the-art methods under both centralized and distributed settings including independent and identically distributed (IID) and non-IID data distributions.

eess.SP↗
arXiv2025

Quantum Properties of Non-Dirichlet Boundary Conditions in Gravity

The Euclidean path integral for gravity is enriched by the addition of boundaries, which provide useful probes of thermodynamic properties. Common boundary conditions include Dirichlet conditions on the boundary induced metric; microcanonical conditions, which refers to fixing some components of the Brown-York boundary stress tensor; and conformal conditions, in which the conformal structure of the induced metric and the trace of the extrinsic curvature are fixed. Boundaries also present interesting problems of consistency. The Dirichlet problem is known, under various (and generally different) conditions, to be inconsistent with perturbative quantization of graviton fluctuations; to exhibit thermodynamic instability; or to require infinite fine-tuning in the presence of matter fluctuations. We extend some of these results to other boundary conditions. We find that similarly to the Dirichlet problem, the graviton fluctuation operator is not elliptic with microcanonical boundaries, and the non-elliptic modes correspond to ``boundary-moving diffeomorphisms." However, we argue that microcanonical factorization of path integrals -- essentially, the insertion of microcanonical constraints on two-sided surfaces in the bulk -- is not affected by the same issues of ellipticity. We also show that for a variety of matter field boundary conditions, matter fluctuations renormalize the gravitational bulk and boundary terms differently, so that the classical microcanonical or conformal variational problems are not preserved unless an infinite fine-tuning is performed.

hep-th↗
arXiv2025

Subteaming and Adaptive Formation Control for Coordinated Multi-Robot Navigation

Coordinated multi-robot navigation is essential for robots to operate as a team in diverse environments. During navigation, robot teams usually need to maintain specific formations, such as circular formations to protect human teammates at the center. However, in complex scenarios such as narrow corridors, rigidly preserving predefined formations can become infeasible. Therefore, robot teams must be capable of dynamically splitting into smaller subteams and adaptively controlling the subteams to navigate through such scenarios while preserving formations. To enable this capability, we introduce a novel method for SubTeaming and Adaptive Formation (STAF), which is built upon a unified hierarchical learning framework: (1) high-level deep graph cut for team splitting, (2) intermediate-level graph learning for facilitating coordinated navigation among subteams, and (3) low-level policy learning for controlling individual mobile robots to reach their goal positions while avoiding collisions. To evaluate STAF, we conducted extensive experiments in both indoor and outdoor environments using robotics simulations and physical robot teams. Experimental results show that STAF enables the novel capability for subteaming and adaptive formation control, and achieves promising performance in coordinated multi-robot navigation through challenging scenarios. More details are available on the project website: https://hcrlab.gitlab.io/project/STAF.

cs.RO↗
arXiv2025

Supersonic gas curtain based real-time ionization profile monitor for hadron therapy

Accurate control and monitoring of the beam is essential for precise dose delivery to tumor tissues during radiotherapy. Real-time monitoring of ion beam profiles and positions improves beam control, patient safety, and treatment reliability by providing immediate feedback. This becomes even more critical in FLASH therapy, where the short corrective window during high-dose delivery demands precise beam control. Existing devices are often limited to in vitro calibration or focus on monitoring a single parameter during treatment. This study aims to develop a device that can simultaneously monitor beam position, profile, current, and energy in real-time, without perturbing the beam, using a supersonic gas curtain system. A supersonic gas curtain beam profile monitor was developed at the Cockcroft Institute to assess its performance and suitability for applications in hadron-beam therapy. The system was integrated with one of the beamlines of the Pelletron accelerator at the Dalton Cumbrian Facility, UK and 2D profile measurements of carbon beams were conducted. The monitor successfully measured the beam profiles within 100ms to 1s across various beam currents (1 - 100 nA), energies (12 - 24 MeV), and charge states (2-5) of carbon. Recorded data was used to estimate detector performance by introducing a parameter called detection limit to quantify sensitivity of the monitor, identifying the threshold number of ions required for detection onset. A method to quantify sensitivity under different beam conditions is then discussed in detail, illustrated with an example case of FLASH beam parameters. This proof-of-concept study demonstrates the performance of the gas curtain-based ionization profile monitor for 2D transverse beam profile measurement of carbon ions. The sensitivity is quantified and evaluated against an example case for FLASH conditions.

physics.acc-ph↗
arXiv2025

A Novel Task-Driven Diffusion-Based Policy with Affordance Learning for Generalizable Manipulation of Articulated Objects

Despite recent advances in dexterous manipulations, the manipulation of articulated objects and generalization across different categories remain significant challenges. To address these issues, we introduce DART, a novel framework that enhances a diffusion-based policy with affordance learning and linear temporal logic (LTL) representations to improve the learning efficiency and generalizability of articulated dexterous manipulation. Specifically, DART leverages LTL to understand task semantics and affordance learning to identify optimal interaction points. The {diffusion-based policy} then generalizes these interactions across various categories. Additionally, we exploit an optimization method based on interaction data to refine actions, overcoming the limitations of traditional diffusion policies that typically rely on offline reinforcement learning or learning from demonstrations. Experimental results demonstrate that DART outperforms most existing methods in manipulation ability, generalization performance, transfer reasoning, and robustness. For more information, visit our project website at: https://sites.google.com/view/dart0257/.

cs.RO↗