Zirui Pang (庞梓瑞)

Zirui Pang (庞梓瑞)

I am currently a first-year PhD student (from Spring 2026) at the Hong Kong University of Science and Technology (Guangzhou) (HKUST-GZ), within the DSA Thrust, Info Hub. I conduct my research under the supervision of Zixin Zhong (钟梓昕) and Jiaheng Wei (魏嘉珩).

My current research interests revolve around:

  1. Noisy Labels: Investigating methods to identify noisy labels (incorrect or missing labels) in various datasets, including images, text, and audio.
  2. LLM Unlearning: Exploring techniques that enable large language models to selectively “forget” specific learned information.
  3. Reinforcement Learning Applications: Studying the application of reinforcement learning in control problems and large-model fine-tuning.
  4. Multi-Agent Systems: Exploring multi-agent systems composed of large language models and their applications in complex task settings.

I have a deep passion for music and enjoy the pleasure it brings me in my leisure time. I previously had a long-term collaboration with the Choir of Central University of Finance and Economics.

📝 Publications

🤖 LLM Unlearning

under review
sym

Label Smoothing Improves Gradient Ascent in LLM Unlearning

Zirui Pang, Hao Zheng, Zhijie Deng, Ling Li, Zixin Zhong, Jiaheng Wei

  • We identify the instability of Gradient Ascent in LLM unlearning.
  • We propose Smoothed Gradient Ascent (SGA) with a tunable smoothing rate.
  • SGA achieves more stable and effective unlearning across benchmarks.
under review
sym

OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models

Hao Zheng, Zirui Pang, Ling li, Zhijie Deng, Yuhan Pu, Zhaowei Zhu, Xiaobo Xia, Jiaheng Wei

Project

  • We present OFFSIDE, a benchmark for multimodal unlearning based on football transfer rumors.
  • It provides real-world, manually curated data and four evaluation settings to test forgetting, utility, and robustness.
  • Our results reveal that current methods fail to unlearn visual rumors and are vulnerable to recovery and prompt attacks.
under review
sym

GUARD: Generation-time LLM Unlearning via Adaptive Restriction and Detection

Zhijie Deng, Chris Yuhao Liu, Zirui Pang, Xinlei He, Lei Feng, Qi Xuan, Zhaowei Zhu, Jiaheng Wei

  • We propose GUARD, a generation-time unlearning framework for LLMs.
  • It detects forget-related prompts and blocks forbidden tokens during generation.
  • GUARD achieves effective forgetting without harming model utility.

🏞️ Label Noise

under review
sym

When VLMs Meet Image Classification: Test Sets Renovation via Missing Label Identification

Zirui Pang, Haosheng Tan, Yuhan Pu, Zhijie Deng, Zhouan Shen, Keyu Hu, Jiaheng Wei

Project

  • We propose REVEAL, a framework that uses vision-language models to find and fix missing and noisy labels in image classification benchmarks.
  • It ensembles multiple VLMs and human feedback to renovate test sets with soft, accurate labels.
  • REVEAL greatly improves dataset quality and aligns closely with human judgments.

📖 Educations

  • 2026.01 - now, PhD Student, Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.
  • 2024.08 - 2025.12, Master, University of Illinois Urbana-Champaign, IL, USA.
  • 2019.09 - 2023.06, Undergraduate, Central University of Finance and Economics, Beijing, China.

💻 Internships

  • 2022.02 - 2022.04, Assistant Software Development Engineer, Iflytek, China.
  • 2024.12 - 2025.12, Research Assistant, HKUST-GZ, China.

🎵 Music

During my undergraduate studies, I maintained a long-term collaboration with the Central University of Finance and Economics Choir. Serving as both the choir’s piano accompanist and a tenor vocalist, I also composed several musical works.

🌍 Visitor Map