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.

Please see my CV and Google Scholar for further details, or contact me at mqiwei@eduhk.hk.

🔬 Research Experience

ET&S 2026
Architecture of the multi-agent conversational AI system for EFL speaking practice

Multi-Agent Conversational AI for EFL Speaking Practice

Jan 2025 - Jan 2026 · Supervisor: Dr. Zhang (SZU)

Jun Zhang, Qiwei Ma, Yu Zhang, and Xiaoming Cao

[Paper] [Code]

  • Designed a conversational system with a lightweight digital-human interface and seven specialised agents for adaptive EFL speaking practice.
  • Evaluated the system in a four-week controlled study with 32 university learners. MAS produced greater gains in oral proficiency (p = 0.049) and grammatical accuracy (p = 0.016) than SAS.
  • MAS supported 26% more practice sessions, 15% longer utterances, and 70% fewer 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

  • Wang, M., Wang, Y., & Ma, Q. (2026). Agentic AI unleashed: Power, promise, and ethical frontiers. In K. T. Chui, C. Li, J. Jia, S. K. S. Cheung, F. W.-F. Lee, & J. Shang (Eds.), Blended learning: Innovations for future education (pp. 27-39). Springer. https://doi.org/10.1007/978-981-92-3761-6_3

📖 Education

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

💻 Experience