CV

Collaborator

Long-Horizon Agent Trajectory Data-Synthesis Pipeline

DeepWisdom

• Surveyed and compared mainstream long-horizon agent benchmarks, distilling their trajectory data-synthesis schemes and their test-task complexity along steps, context, tools, and environment into a comparative analysis (covering 12+ synthetic-trajectory works)
• Identified the limitations of existing benchmarks in data scale and real-execution environments, providing design guidance for how the pipeline should generate longer-horizon, more realistic trajectory data

Developer

High-Quality Image-Text Data Synthesis · Image-Caption Evaluation Agent

Prof. Jiaheng Wei’s Group · Huawei Collaboration (ongoing)

• First to apply an agent system to this task: designed and implemented a caption-precision evaluation scheme in which the agent understands an image hierarchically, from coarse to fine and aligned with human perception, and reconstructs a scene graph — enabling reference-free caption evaluation
• Aligns caption units (objects, attributes, relations) with the hierarchical scene graph to judge whether a caption precisely and completely describes the image; the hierarchical decomposition and scene-graph reconstruction work well and reliably filter out hallucinated and attribute-mismatched captions

BEng

The Hong Kong University of Science and Technology (Guangzhou)

Artificial Intelligence

• CGPA: 3.6+/4.3 (top 15%)

  Machine Learning: A (top 15%)

  Ethics, Privacy & Security in AI: A+

Exchange Student

RWTH Aachen University

Exchange Program

• Deep Learning (full marks)

IPV-Bench: Benchmarking Image Protection Methods under Diverse Image-to-Video Generation Scenarios

arXiv:2603.26154 (under review at AAAI 2027)

The first systematic benchmark for image protection methods in image-to-video (I2V) generation scenarios — unifying evaluation protocols and generation settings, covering fidelity–effectiveness trade-offs, cross-model transfer, and robustness across diverse image domains. † Corresponding author.

Programming Guide for Solving Constraint Satisfaction Problems with Tensor Networks

arXiv:2501.00227

A practical programming guide demonstrating how to formulate and solve constraint satisfaction problems (e.g., graph coloring, SAT) with tensor-network-based methods, with reproducible Julia implementations. † Corresponding author.

Third Prize, Huawei Ascend Operator Challenge (S8)

Merit Award, Undergraduate Research Project Competition

HKUST(GZ)

Dean's List

HKUST(GZ)