Qiwei Ma(马启伟)

Qiwei Ma(马启伟)

Research Assistant at The Education University of Hong Kong

I’m currently a Research Assistant at The Education University of Hong Kong. My current research focuses on Agentic AI, Embodied Intelligence, and AI in Education.

I will be applying for PhD positions for Fall 2027 entry and am also open to earlier opportunities after completing my RA contract.

Download my CV here.

🔬 Research Experience

Multi-Agent Conversational AI for EFL Speaking Practice (Jan 2025 - Jan 2026) — Supervisor: Dr. Zhang (SZU)

  • Developed a multi-agent system (MAS) for EFL speaking practice.
  • Built seven specialized agents for preprocessing, response generation, and dialogue supervision.
  • Investigated the mechanisms underlying MAS superiority and identified synergistic effects among integrated features.
  • Conducted a 4-week controlled experiment with 32 university EFL learners, showing that MAS outperformed SAS in oral proficiency gains (p = 0.049) and grammatical accuracy (p = 0.016).
  • Found that MAS led to 26% more practice sessions, 15% longer utterances, and a 70% reduction in repeated grammatical errors.

Reasoning for Table Manipulation (Mar 2025 - Jan 2026) — Supervisors: Dr. Yang, Dr. Tan (SIAT, CAS)

  • Developed an end-to-end LLM framework for table manipulation through structured reasoning.
  • Constructed a benchmark covering five core tasks: table splitting and merging, wide-to-long conversion, semi-structured field parsing, and row/column generation.
  • Completed two-stage training with supervised fine-tuning on reasoning traces followed by GRPO optimization, achieving state-of-the-art performance among 7B-scale table-specific models.
  • Analyzed structural integrity challenges and found that column-level accuracy often exceeded row-level accuracy because of sensitivity to missing fields.

🔧 Additional Research

Differential Privacy Image Generation (Mar 2024 - Apr 2025) — Supervisor: Dr. Zhang (SZU)

  • Developed a differential privacy framework for image generation to reduce gradient clipping bias and improve training stability.
  • Introduced reconstruction loss and noise injection during generator upsampling to improve data utility and image diversity.
  • Designed a multi-component training pipeline integrating a generator, discriminators, a classifier, and an encoder with gradient sanitization mechanisms.
  • Achieved state-of-the-art performance on MNIST and Fashion-MNIST, surpassing baseline methods in Inception Score, Frechet Inception Distance, and downstream classification accuracy.

Intelligent Annotation and Feedback System for English Writing (Apr 2025 - Sep 2025) — Supervisor: Dr. Zhang (SZU)

  • Developed LLM-IAF (LLM-based Intelligent Annotation and Feedback System), a mobile application for automated English writing assessment for junior high school students.
  • Implemented a dual-engine AI workflow combining semantic evaluation and visual grounding to provide immediate visual feedback with error localization on handwritten essays.
  • Conducted an experimental study comparing students using LLM-IAF with a control group receiving traditional instruction across four weekly writing tasks.
  • Observed significant improvements in writing performance, learning engagement, and writing self-efficacy in the experimental group.
  • Demonstrated strong agreement between AI and teacher scoring, and students reported high satisfaction with the system’s usefulness and ease of use.

📝 Publications

  • Zhang, J., Ma, Q., Zhang, Y., & Cao, X. (2026). Multi-agent vs. single-agent AI for EFL speaking practice: A controlled experiment with hybrid input, contextual dialogue, and proficiency-adaptive feedback. Educational Technology & Society, 29(2), 297-322. https://doi.org/10.30191/ETS.202604_29(2).SP05

📖 Education

  • 2023.09 - 2026.06, Master, Shenzhen University, Shenzhen.
  • 2018.09 - 2022.06, Undergraduate, Ningxia University, Yinchuan.

💻 Experience