CV

Education

• M.S. in Human-Centered Artificial Intelligent, Technical University of Denmark, 2023-2025 GPA:3.34/4

Experience

• 2025.4.1 - 2025.8.31: Research Assistant
  • MBZUAI’s Research Assistant, supervised by Prof. Yova Kementchedjhieva.
• 2025.1.06 - 2025.3.31: Visiting Researcher
  • MBZUAI’s Visiting Researcher, supervised by Prof. Yova Kementchedjhieva.
• 2024: Research Assistant（Remote）
  • Collabrate with Shanghai AI Laboratory and Xiamen University in DViR: Dynamic Visual Routing for Weakly Supervised Referring
• 2023: Research Assistant（Remote）
  • Collabrate with Xia Men University and Peng Cheng Laboratory in Weakly Supervised Referring Expression Comphrehension
• November 2023: Conference Helper
  • 30th Annual Conference on Computer and Communications Security(CCS)

Expression Comprehension

Projects

• HOW DO GUI AGENTS FAIL? A TRAJECTORY-LEVEL BEHAVIORAL ANALYSIS STUDY. This project studies GUI-agent failure through full interaction trajectories rather than final accuracy alone, using 360 trajectories from six multimodal models across six CAPTCHA-style puzzle families. Key points:
  • Proposed trajectory-level behavioral signatures to reveal hidden differences in interaction cost, tool use, repeated loops, and stopping behavior.
  • Analyzed programmatic evaluate calls and separated GUI assistance from event-level or state-level GUI bypass.
  • Identified distinct failure regimes, including early stopping, long unproductive interaction, over-probing, and self-correction without recovery.
• FLASH-SAM: KNOWLEDGE DISTILLATION FOR SEGMENT ANYTHING (ADVANCED COMPUTER VISION COURSE PROJECT). By experimenting with feature-based and logits-based distillation methods for the SAM model’s image encoder, and ultimately using the feature-based approach along with a newly proposed feature adapter, the results surpassed those of MobileSAM. Roles in charge:
  • Collect, read and analysis the paper from this research area and discuss advantages and disadvantages of each pathway to find novel ideas with supervisor
  • Implement and improve the research ideas and conduct extensive experiments to prove it work.
  • Write the research paper, make the project poster and present the presentation.
• ENHANCE WEAKLY SUPERVISED REFERRING IMAGE SEGMENTATION WITH DIFFUSION MODEL (ADVISED BY PROF. DIMITRIOS PAPADOPOULOS AT DTU). The model uses attention maps from a diffusion model to enhance image segmentation by focusing on relevant features based on linguistic cues. It employs a pyramid CNN architecture to integrate attention maps at multiple scales, improving segmentation accuracy and computational efficiency for weakly supervised RES applications. Roles in charge:
  • Collect, read and analysis the paper from this research area and discuss advantages and disadvantages of each pathway to find novel ideas with supervisor
  • Implement and improve the research ideas and conduct extensive experiments to prove it work.
  • Write the research paper, make the project poster and present the presentation.