Heng Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Heng Zhang has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- Jin Huang
South China Normal University, Guangzhou, China
2 shared publications - Dandan Zhu
College of Information Science and Engineering, China University of Petroleum (Beijing), Beijing, China
1 shared publication - Hangxiangpan Wang
China University of Petroleum (Beijing), Beijing, China and Hainan Institute of China University of Petroleum (Beijing), Sanya, Hainan, China
1 shared publication - Hao Zhang
University of Chinese Academy of Sciences, Beijing, China
1 shared publication - HuiShen Jiao
China University of Petroleum (Beijing), Beijing, China
1 shared publication - Jiakai Pan
School of Microelectronics Science and Technology, Sun Yat-Sen University, Zhuhai, China
1 shared publication
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 14350-14359
HyperBalance: Hypergraph-based Topology Optimization for Reliable Multi-Agent Communication
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 14350-14359
Open AccessHyperBalance: Hypergraph-based Topology Optimization for Reliable Multi-Agent Communication
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1851-1861
H4G: Unlocking Faithful Inference for Zero-Shot Graph Learning in Hyperbolic Space
Text-attributed graphs are widely used across domains, offering rich opportunities for zero-shot learning via graph-text alignment. However, existing methods struggle with tasks requiring fine-grained pattern recognition, particularly on heterophilic graphs. Through empirical and theoretical analysis, we identify an over-abstraction problem: current approaches operate at excessively large hyperbolic radii, compressing multi-scale structural information into uniform high-level abstractions. This abstraction-induced information loss obscures critical local patterns essential for accurate predictions. By analyzing embeddings in hyperbolic space, we demonstrate that optimal graph learning requires faithful preservation of fine-grained structural details, better retained by representations positioned closer to the origin. To address this, we propose H4G, a framework that systematically reduces embedding radii using learnable block-diagonal scaling matrices and Möbius matrix multiplication. This approach restores access to fine-grained patterns while maintaining global receptive ability with minimal computational overhead. Experiments show H4G achieves state-of-the-art zero-shot performance with 12.8% improvement on heterophilic graphs and 8.4% on homophilic graphs, confirming that radius reduction enables faithful multi-scale representation for advancing zero-shot graph learning.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1851-1861
Open AccessH4G: Unlocking Faithful Inference for Zero-Shot Graph Learning in Hyperbolic Space
Text-attributed graphs are widely used across domains, offering rich opportunities for zero-shot learning via graph-text alignment. However, existing methods struggle with tasks requiring fine-grained pattern recognition, particularly on heterophilic graphs. Through empirical and theoretical analysis, we identify an over-abstraction problem: current approaches operate at excessively large hyperbolic radii, compressing multi-scale structural information into uniform high-level abstractions. This abstraction-induced information loss obscures critical local patterns essential for accurate predictions. By analyzing embeddings in hyperbolic space, we demonstrate that optimal graph learning requires faithful preservation of fine-grained structural details, better retained by representations positioned closer to the origin. To address this, we propose H4G, a framework that systematically reduces embedding radii using learnable block-diagonal scaling matrices and Möbius matrix multiplication. This approach restores access to fine-grained patterns while maintaining global receptive ability with minimal computational overhead. Experiments show H4G achieves state-of-the-art zero-shot performance with 12.8% improvement on heterophilic graphs and 8.4% on homophilic graphs, confirming that radius reduction enables faithful multi-scale representation for advancing zero-shot graph learning.
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 525-534
Hyperbolic and Evidence-Prioritized Experts for Large Vision-Language Models
- 15 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 525-534
Open AccessHyperbolic and Evidence-Prioritized Experts for Large Vision-Language Models
- Open Access
- 2 February 2026
- Published by Elsevier in SSRN Electronic Journal
Shifting the Computational Burden: Efficient Long-Text MLLMs via LLM-Guided Visual Text Rendering
To address the significant computational overhead and semantic alignment challenges of processing long texts in Multimodal Large Models (MLLMs), this paper introduces a novel framework named \textbf{VisCompConText}. This framework leverages two key innovations—-visual text compression and LLM-guided training—to efficiently extend the context-handling capabilities of multimodal models. The core mechanism involves rendering lengthy text segments into spatially compact visual images (i.e., visual tokens), which drastically reduces GPU memory consumption and computational load during both training and inference. Concurrently, we leverage the superior comprehension abilities of Large Language Models (LLMs) to guide the training of the visual encoder, significantly enhancing its deep semantic awareness of long-form text. Experiments demonstrate that \textbf{VisCompConText} effectively improves model performance on tasks such as long-document understanding and visual question answering, complementing existing context-extension techniques. This work provides a systematic solution for achieving long-context multimodal learning that balances computational efficiency with semantic fidelity.
- 2 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessShifting the Computational Burden: Efficient Long-Text MLLMs via LLM-Guided Visual Text Rendering
To address the significant computational overhead and semantic alignment challenges of processing long texts in Multimodal Large Models (MLLMs), this paper introduces a novel framework named \textbf{VisCompConText}. This framework leverages two key innovations—-visual text compression and LLM-guided training—to efficiently extend the context-handling capabilities of multimodal models. The core mechanism involves rendering lengthy text segments into spatially compact visual images (i.e., visual tokens), which drastically reduces GPU memory consumption and computational load during both training and inference. Concurrently, we leverage the superior comprehension abilities of Large Language Models (LLMs) to guide the training of the visual encoder, significantly enhancing its deep semantic awareness of long-form text. Experiments demonstrate that \textbf{VisCompConText} effectively improves model performance on tasks such as long-document understanding and visual question answering, complementing existing context-extension techniques. This work provides a systematic solution for achieving long-context multimodal learning that balances computational efficiency with semantic fidelity.