arXiv subjects

Dong Zhang

Publications and source records attributed to Dong Zhang.

At least 73 records · Page 4Linked to original sources

arXiv2025

Nova: Real-Time Agentic Vision-Language Model Serving with Adaptive Cross-Stage Parallelization

This paper presents Nova, a real-time scheduling framework for serving agentic vision-language models (VLMs) on a single GPU with balanced per-request latency and overall request process throughput. Our design begins by enabling effective pipelining across vision encode, LLM prefill, and LLM decode stages of VLMs, by exploiting their heterogeneous resource demands during execution and incorporating elastic GPU spatial partitioning among stages to maximally utilize the compute and memory resources. Building on this, we introduce a real-time scheduling algorithm that adaptively calibrates resource allocation among stages based on a Pareto-optimal analysis of the latency-throughput trade-off, allowing the system to sustain responsiveness and resource efficiency under dynamic request loads. To further alleviate GPU memory pressure, we design a lightweight weight offloading strategy for vision encoders that preserves inference efficiency with minimized memory overhead. Extensive evaluations on both synthetic and real-world agent workloads demonstrate that Nova consistently outperforms the state-of-the-art baselines, improving the maximum latency by up to 23.3%, while keeping competitive throughput.

cs.OS↗

arXiv2025

VStyle: A Benchmark for Voice Style Adaptation with Spoken Instructions

Spoken language models (SLMs) have emerged as a unified paradigm for speech understanding and generation, enabling natural human machine interaction. However, while most progress has focused on semantic accuracy and instruction following, the ability of SLMs to adapt their speaking style based on spoken instructions has received limited attention. We introduce Voice Style Adaptation (VSA), a new task that examines whether SLMs can modify their speaking style, such as timbre, prosody, or persona following natural language spoken commands. To study this task, we present VStyle, a bilingual (Chinese & English) benchmark covering four categories of speech generation: acoustic attributes, natural language instruction, role play, and implicit empathy. We also introduce the Large Audio Language Model as a Judge (LALM as a Judge) framework, which progressively evaluates outputs along textual faithfulness, style adherence, and naturalness, ensuring reproducible and objective assessment. Experiments on commercial systems and open source SLMs demonstrate that current models face clear limitations in controllable style adaptation, highlighting both the novelty and challenge of this task. By releasing VStyle and its evaluation toolkit, we aim to provide the community with a foundation for advancing human centered spoken interaction. The dataset and code are publicly available at \href{https://junzhan2000.github.io/VStyle.github.io/}{project's homepage}.

cs.SD↗

arXiv2025

A.S.E: A Repository-Level Benchmark for Evaluating Security in AI-Generated Code

The increasing adoption of large language models (LLMs) in software engineering necessitates rigorous security evaluation of their generated code. However, existing benchmarks often lack relevance to real-world AI-assisted programming scenarios, making them inadequate for assessing the practical security risks associated with AI-generated code in production environments. To address this gap, we introduce A.S.E (AI Code Generation Security Evaluation), a repository-level evaluation benchmark designed to closely mirror real-world AI programming tasks, offering a comprehensive and reliable framework for assessing the security of AI-generated code. Our evaluation of leading LLMs on A.S.E reveals several key findings. In particular, current LLMs still struggle with secure coding. The complexity in repository-level scenarios presents challenges for LLMs that typically perform well on snippet-level tasks. Moreover, a larger reasoning budget does not necessarily lead to better code generation. These observations offer valuable insights into the current state of AI code generation and help developers identify the most suitable models for practical tasks. They also lay the groundwork for refining LLMs to generate secure and efficient code in real-world applications.

cs.SE↗

arXiv2025

Decoupled Proxy Alignment: Mitigating Language Prior Conflict for Multimodal Alignment in MLLM

Multimodal large language models (MLLMs) have gained significant attention due to their impressive ability to integrate vision and language modalities. Recent advancements in MLLMs have primarily focused on improving performance through high-quality datasets, novel architectures, and optimized training strategies. However, in this paper, we identify a previously overlooked issue, language prior conflict, a mismatch between the inherent language priors of large language models (LLMs) and the language priors in training datasets. This conflict leads to suboptimal vision-language alignment, as MLLMs are prone to adapting to the language style of training samples. To address this issue, we propose a novel training method called Decoupled Proxy Alignment (DPA). DPA introduces two key innovations: (1) the use of a proxy LLM during pretraining to decouple the vision-language alignment process from language prior interference, and (2) dynamic loss adjustment based on visual relevance to strengthen optimization signals for visually relevant tokens. Extensive experiments demonstrate that DPA significantly mitigates the language prior conflict, achieving superior alignment performance across diverse datasets, model families, and scales. Our method not only improves the effectiveness of MLLM training but also shows exceptional generalization capabilities, making it a robust approach for vision-language alignment. Our code is available at https://github.com/fnlp-vision/DPA.

cs.CL↗

arXiv2025

UnifiedVisual: A Framework for Constructing Unified Vision-Language Datasets

Unified vision large language models (VLLMs) have recently achieved impressive advancements in both multimodal understanding and generation, powering applications such as visual question answering and text-guided image synthesis. However, progress in unified VLLMs remains constrained by the lack of datasets that fully exploit the synergistic potential between these two core abilities. Existing datasets typically address understanding and generation in isolation, thereby limiting the performance of unified VLLMs. To bridge this critical gap, we introduce a novel dataset construction framework, UnifiedVisual, and present UnifiedVisual-240K, a high-quality dataset meticulously designed to facilitate mutual enhancement between multimodal understanding and generation. UnifiedVisual-240K seamlessly integrates diverse visual and textual inputs and outputs, enabling comprehensive cross-modal reasoning and precise text-to-image alignment. Our dataset encompasses a wide spectrum of tasks and data sources, ensuring rich diversity and addressing key shortcomings of prior resources. Extensive experiments demonstrate that models trained on UnifiedVisual-240K consistently achieve strong performance across a wide range of tasks. Notably, these models exhibit significant mutual reinforcement between multimodal understanding and generation, further validating the effectiveness of our framework and dataset. We believe UnifiedVisual represents a new growth point for advancing unified VLLMs and unlocking their full potential. Our code and datasets is available at https://github.com/fnlp-vision/UnifiedVisual.

cs.CL↗

arXiv2025

Mitigating Query Selection Bias in Referring Video Object Segmentation

Recently, query-based methods have achieved remarkable performance in Referring Video Object Segmentation (RVOS) by using textual static object queries to drive cross-modal alignment. However, these static queries are easily misled by distractors with similar appearance or motion, resulting in \emph{query selection bias}. To address this issue, we propose Triple Query Former (TQF), which factorizes the referring query into three specialized components: an appearance query for static attributes, an intra-frame interaction query for spatial relations, and an inter-frame motion query for temporal association. Instead of relying solely on textual embeddings, our queries are dynamically constructed by integrating both linguistic cues and visual guidance. Furthermore, we introduce two motion-aware aggregation modules that enhance object token representations: Intra-frame Interaction Aggregation incorporates position-aware interactions among objects within a single frame, while Inter-frame Motion Aggregation leverages trajectory-guided alignment across frames to ensure temporal coherence. Extensive experiments on multiple RVOS benchmarks demonstrate the advantages of TQF and the effectiveness of our structured query design and motion-aware aggregation modules.

cs.CV↗

arXiv2025

Resource-Efficient Glioma Segmentation on Sub-Saharan MRI

Gliomas are the most prevalent type of primary brain tumors, and their accurate segmentation from MRI is critical for diagnosis, treatment planning, and longitudinal monitoring. However, the scarcity of high-quality annotated imaging data in Sub-Saharan Africa (SSA) poses a significant challenge for deploying advanced segmentation models in clinical workflows. This study introduces a robust and computationally efficient deep learning framework tailored for resource-constrained settings. We leveraged a 3D Attention UNet architecture augmented with residual blocks and enhanced through transfer learning from pre-trained weights on the BraTS 2021 dataset. Our model was evaluated on 95 MRI cases from the BraTS-Africa dataset, a benchmark for glioma segmentation in SSA MRI data. Despite the limited data quality and quantity, our approach achieved Dice scores of 0.76 for the Enhancing Tumor (ET), 0.80 for Necrotic and Non-Enhancing Tumor Core (NETC), and 0.85 for Surrounding Non-Functional Hemisphere (SNFH). These results demonstrate the generalizability of the proposed model and its potential to support clinical decision making in low-resource settings. The compact architecture, approximately 90 MB, and sub-minute per-volume inference time on consumer-grade hardware further underscore its practicality for deployment in SSA health systems. This work contributes toward closing the gap in equitable AI for global health by empowering underserved regions with high-performing and accessible medical imaging solutions.

cs.CV↗

arXiv2025

AnyGPT: Unified Multimodal LLM with Discrete Sequence Modeling

We introduce AnyGPT, an any-to-any multimodal language model that utilizes discrete representations for the unified processing of various modalities, including speech, text, images, and music. AnyGPT can be trained stably without any alterations to the current large language model (LLM) architecture or training paradigms. Instead, it relies exclusively on data-level preprocessing, facilitating the seamless integration of new modalities into LLMs, akin to the incorporation of new languages. We build a multimodal text-centric dataset for multimodal alignment pre-training. Utilizing generative models, we synthesize the first large-scale any-to-any multimodal instruction dataset. It consists of 108k samples of multi-turn conversations that intricately interweave various modalities, thus equipping the model to handle arbitrary combinations of multimodal inputs and outputs. Experimental results demonstrate that AnyGPT is capable of facilitating any-to-any multimodal conversation while achieving performance comparable to specialized models across all modalities, proving that discrete representations can effectively and conveniently unify multiple modalities within a language model. Demos are shown in https://junzhan2000.github.io/AnyGPT.github.io/

cs.CL↗

arXiv2025

Zero-shot Cross-lingual NER via Mitigating Language Difference: An Entity-aligned Translation Perspective

Cross-lingual Named Entity Recognition (CL-NER) aims to transfer knowledge from high-resource languages to low-resource languages. However, existing zero-shot CL-NER (ZCL-NER) approaches primarily focus on Latin script language (LSL), where shared linguistic features facilitate effective knowledge transfer. In contrast, for non-Latin script language (NSL), such as Chinese and Japanese, performance often degrades due to deep structural differences. To address these challenges, we propose an entity-aligned translation (EAT) approach. Leveraging large language models (LLMs), EAT employs a dual-translation strategy to align entities between NSL and English. In addition, we fine-tune LLMs using multilingual Wikipedia data to enhance the entity alignment from source to target languages.

cs.CL↗

arXiv2025

Notes on su$(1,2)\oplus$u$(1)$ Chern-Simons theory and Torsional Newton-Cartan gravity

In this study, we investigate three-dimensional torsional Newton-Cartan (TNC) gravity by gauging the su$(1,2)\oplus$u$(1)$ algebra and construct its action using the Chern-Simons theory. This TNC exhibits novel features, including the fact that the gauge fields associated with both dilatation and rotation symmetries transform non-trivially under Galilean boosts. This theory also reproduces the Schrödinger gravity acquired by gauging the extended $z=2$ Schrödinger algebra arXiv:1604.08054 via a large speed of light ($1/c$)-expansion. In particular, we explain that the $z=2$ Lifshitz vacuum solution appearing in Schrödinger gravity is related to the null reduction of 4d $Ω$-background up to a conformal factor. Based on these results, we revisit the identification between the extended Schrödinger algebra and bosonic analogue of super BMS algebra arXiv:1905.13154. We interpret that this relation originates from the $\mathcal{W}_3^{(2)}$ algebra which acts as the bosonic analogue of $\mathcal{N}=2$ superconformal algebra.

hep-th↗

arXiv2025

Towards Customized Knowledge Distillation for Chip-Level Dense Image Predictions

It has been revealed that efficient dense image prediction (EDIP) models designed for AI chips, trained using the knowledge distillation (KD) framework, encounter two key challenges, including \emph{maintaining boundary region completeness} and \emph{ensuring target region connectivity}, despite their favorable real-time capacity to recognize the main object regions. In this work, we propose a customized boundary and context knowledge distillation (BCKD) method for EDIPs, which facilitates the targeted KD from large accurate teacher models to compact small student models. Specifically, the \emph{boundary distillation} focuses on extracting explicit object-level boundaries from the hierarchical feature maps to enhance the student model's mask quality in boundary regions. Meanwhile, the \emph{context distillation} leverages self-relations as a bridge to transfer implicit pixel-level contexts from the teacher model to the student model, ensuring strong connectivity in target regions. Our proposed method is specifically designed for the EDIP tasks and is characterized by its simplicity and efficiency. Theoretical analysis and extensive experimental results across semantic segmentation, object detection, and instance segmentation on five representative datasets demonstrate the effectiveness of BCKD, resulting in well-defined object boundaries and smooth connecting regions.

cs.CV↗

arXiv2025

Anomalous magnetoresistance in an antiferromagnetic Kagome semimetal heterostructures

Antiferromagnetic Kagome semimetals have attracted tremendous attentions for their potential application in antiferromagnetic topological spintronics. Effectively manipulating Kagome antiferromagnetic states could reveal abundant physical phenomena induced from quantum interactions between topology, spin, and correlation. Here, we achieved tunable spin textures of FeSn thin films via introducing interfacial Dzyaloshinskii Moriya interaction from heavy-metal Pt overlayer. With increasing FeSn thickness, the variable spin textures result in gradual change in Hall resistivity and magnetoresistance. Importantly, an unconventional damped oscillatory-like behavior of magnetoresistance at relatively low magnetic field can be observed in thin FeSn-Pt samples. This oscillatory like magnetoresistance feature was confirmed to be related to the special topological spin textures revealed by magnetic force microscopy measurements. The formation of rich variety of topological spin textures in association with exotic magneto-transport properties in antiferromagnetic Kagome FeSn heterostructures offers new perspectives for understanding the novel emergent phenomena in Kagome antiferromagnets.

cond-mat.mtrl-sci↗

arXiv2025

Diffusion-Guided Knowledge Distillation for Weakly-Supervised Low-Light Semantic Segmentation

Weakly-supervised semantic segmentation aims to assign category labels to each pixel using weak annotations, significantly reducing manual annotation costs. Although existing methods have achieved remarkable progress in well-lit scenarios, their performance significantly degrades in low-light environments due to two fundamental limitations: severe image quality degradation (e.g., low contrast, noise, and color distortion) and the inherent constraints of weak supervision. These factors collectively lead to unreliable class activation maps and semantically ambiguous pseudo-labels, ultimately compromising the model's ability to learn discriminative feature representations. To address these problems, we propose Diffusion-Guided Knowledge Distillation for Weakly-Supervised Low-light Semantic Segmentation (DGKD-WLSS), a novel framework that synergistically combines Diffusion-Guided Knowledge Distillation (DGKD) with Depth-Guided Feature Fusion (DGF2). DGKD aligns normal-light and low-light features via diffusion-based denoising and knowledge distillation, while DGF2 integrates depth maps as illumination-invariant geometric priors to enhance structural feature learning. Extensive experiments demonstrate the effectiveness of DGKD-WLSS, which achieves state-of-the-art performance in weakly supervised semantic segmentation tasks under low-light conditions. The source codes have been released at:https://github.com/ChunyanWang1/DGKD-WLSS.

cs.CV↗

arXiv2025

R-Genie: Reasoning-Guided Generative Image Editing

While recent advances in image editing have enabled impressive visual synthesis capabilities, current methods remain constrained by explicit textual instructions and limited editing operations, lacking deep comprehension of implicit user intentions and contextual reasoning. In this work, we introduce a new image editing paradigm: reasoning-guided generative editing, which synthesizes images based on complex, multi-faceted textual queries accepting world knowledge and intention inference. To facilitate this task, we first construct a comprehensive dataset featuring over 1,000 image-instruction-edit triples that incorporate rich reasoning contexts and real-world knowledge. We then propose R-Genie: a reasoning-guided generative image editor, which synergizes the generation power of diffusion models with advanced reasoning capabilities of multimodal large language models. R-Genie incorporates a reasoning-attention mechanism to bridge linguistic understanding with visual synthesis, enabling it to handle intricate editing requests involving abstract user intentions and contextual reasoning relations. Extensive experimental results validate that R-Genie can equip diffusion models with advanced reasoning-based editing capabilities, unlocking new potentials for intelligent image synthesis.

cs.CV↗

arXiv2025

XY-Tokenizer: Mitigating the Semantic-Acoustic Conflict in Low-Bitrate Speech Codecs

Speech codecs serve as bridges between speech signals and large language models. An ideal codec for speech language models should not only preserve acoustic information but also capture rich semantic information. However, existing speech codecs struggle to balance high-quality audio reconstruction with ease of modeling by language models. In this study, we analyze the limitations of previous codecs in balancing semantic richness and acoustic fidelity. We propose XY-Tokenizer, a novel codec that mitigates the conflict between semantic and acoustic capabilities through multi-stage, multi-task learning. Experimental results demonstrate that XY-Tokenizer achieves performance in both semantic and acoustic tasks comparable to that of state-of-the-art codecs operating at similar bitrates, even though those existing codecs typically excel in only one aspect. Specifically, XY-Tokenizer achieves strong text alignment, surpassing distillation-based semantic modeling methods such as SpeechTokenizer and Mimi, while maintaining a speaker similarity score of 0.83 between reconstructed and original audio. The reconstruction performance of XY-Tokenizer is comparable to that of BigCodec, the current state-of-the-art among acoustic-only codecs, which achieves a speaker similarity score of 0.84 at a similar bitrate. Code and models are available at https://github.com/gyt1145028706/XY-Tokenizer.

cs.SD↗

arXiv2025

A Comprehensive Graph Framework for Question Answering with Mode-Seeking Preference Alignment

Recent advancements in retrieval-augmented generation (RAG) have enhanced large language models in question answering by integrating external knowledge. However, challenges persist in achieving global understanding and aligning responses with human ethical and quality preferences. To address these issues, we propose GraphMPA, a comprehensive graph-based framework with mode-seeking preference alignment. Our approach constructs a hierarchical document graph using a general similarity measurement, mimicking human cognitive processes for information understanding and synthesis. Additionally, we introduce mode-seeking preference optimization to better align model outputs with human preferences through probability-matching constraints. Extensive experiments on six datasets demonstrate the effectiveness of our \href{https://github.com/tangquanwei/GraphMPA}{GraphMPA}.

cs.CL↗

arXiv2025

InstructTTSEval: Benchmarking Complex Natural-Language Instruction Following in Text-to-Speech Systems

In modern speech synthesis, paralinguistic information--such as a speaker's vocal timbre, emotional state, and dynamic prosody--plays a critical role in conveying nuance beyond mere semantics. Traditional Text-to-Speech (TTS) systems rely on fixed style labels or inserting a speech prompt to control these cues, which severely limits flexibility. Recent attempts seek to employ natural-language instructions to modulate paralinguistic features, substantially improving the generalization of instruction-driven TTS models. Although many TTS systems now support customized synthesis via textual description, their actual ability to interpret and execute complex instructions remains largely unexplored. In addition, there is still a shortage of high-quality benchmarks and automated evaluation metrics specifically designed for instruction-based TTS, which hinders accurate assessment and iterative optimization of these models. To address these limitations, we introduce InstructTTSEval, a benchmark for measuring the capability of complex natural-language style control. We introduce three tasks, namely Acoustic-Parameter Specification, Descriptive-Style Directive, and Role-Play, including English and Chinese subsets, each with 1k test cases (6k in total) paired with reference audio. We leverage Gemini as an automatic judge to assess their instruction-following abilities. Our evaluation of accessible instruction-following TTS systems highlights substantial room for further improvement. We anticipate that InstructTTSEval will drive progress toward more powerful, flexible, and accurate instruction-following TTS.

cs.CL↗

arXiv2025

Relaxation-Free Min-k-Partition for PCI Assignment in 5G Networks

Physical Cell Identity (PCI) is a critical parameter in 5G networks. Efficient and accurate PCI assignment is essential for mitigating mod-3 interference, mod-30 interference, collisions, and confusions among cells, which directly affect network reliability and user experience. In this paper, we propose a novel framework for PCI assignment by decomposing the problem into Min-3-Partition, Min-10-Partition, and a graph coloring problem, leveraging the Chinese Remainder Theorem (CRT). Furthermore, we develop a relaxation-free approach to the general Min-k-Partition problem by reformulating it as a quadratic program with a norm-equality constraint and solving it using a penalized mirror descent (PMD) algorithm. The proposed method demonstrates superior computational efficiency and scalability, significantly reducing interference while eliminating collisions and confusions in large-scale 5G networks. Numerical evaluations on real-world datasets show that our approach reduces computational time by up to 20 times compared to state-of-the-art methods, making it highly practical for real-time PCI optimization in large-scale networks. These results highlight the potential of our method to improve network performance and reduce deployment costs in modern 5G systems.

eess.SP↗