Zeyu Qin

Zeyu Qin

Ph.D. student of Computer Science & Engineering at HKUST

πŸŽ“ About Me

I am Zeyu Qin (秦泽钰), a forth-year Ph.D. Candidate of Computer Science & Engineering at Hong Kong University of Science and Technology (HKUST). My advisor is Prof. Minhao Cheng.

I am currently doing research on long-horizon agents at Kimi (Kimi K2 thinking, K2.5). I have also had wonderful internship experiences at MSRA GenAI group, Alibaba DAMO Academy (Now Tongyi Lab), and Tencent AI Lab. My research interests broadly span Long-Horizon Agents, Reasoning, and Scalable Oversight.

Currently, I am focusing on:
1) Long-Horizon Agent: Developing Generalizable Long-Horizon Agent
2) Scaling Training-time and Testing-time Compute: Scaling Compute to Activate Capacity and Enable Fast Adaptation
3) Training Critique Models: Providing Reliable Supervision Signals

I am very fortunate to have worked with lots of distinguished researchers: Dr. Xingxing Zhang, Dr. Liuyi Yao, Daoyuan Chen, Dr. Yaliang Li, Prof. Baoyuan Wu, Dr. Yanbo Fan, Prof. Hongyuan Zha, Dr. Jiancong Xiao, and Prof. Piji Li. I had an amazing four years at LGU with my good friends from Room 310 and Room 224. Miss you, guys!

πŸ“œ News

  • 2026.02：Kimi K2.5 tech report has been released [arxiv].
  • 2025.07：Scaling Laws of Synthetic Data for Language Models has been accepted to COLM 2025.
  • 2025.05: Safety Reasoning paper has been accepted to ICML 2025: we explore the Safety Reasoning with Guidelines. We introduce SRG, a scalable framework that synthesizes safety CoT supervision to encourage models to reason in alignment with diverse safety guidelines, each reflecting a different perspective on safety knowledge. Our SRG significantly improves OOD generalization of safety alignment.
  • 2025.03: New work: we explore the Scaling Laws of Synthetic Data for Language Models. We introducing SynthLLM, a scalable framework that transforms pre-training corpora into diverse, high-quality synthetic datasets and find synthetic data that reliably adheres to the rectified scaling law across various model sizes.
  • 2024.09: One Spotlight has been accepted to NeurIPS 2024: Understanding Superficial safety of Safety Tuning–the Backdoor Purification Case. We demonstrated that safety fine-tuning cannot fully eliminate learned harmful features and provided a detailed analysis.

πŸ“ Publications

* denoting equal contribution $^{\dagger}$ denoting corresponding author

Tech Report

  • Kimi K2. 5: Visual Agentic Intelligence [arxiv]
    Contribute to Long-Horizon Agent

  • Kimi K2: Open Agentic Intelligence [arxiv]
    Contribute to K2 Thinking

Scalable and Reliable Oversight

  • Scaling Laws of Synthetic Data for Language Models
    Zeyu Qin, Qingxiu Dong, Xingxing Zhang $^{\dagger}$, Li Dong, Xiaolong Huang, Ziyi Yang, Mahmoud Khademi, Dongdong Zhang, Hany Hassan Awadalla, Yi R. Fung, Weizhu Chen, Minhao Cheng, Furu Wei $^{\dagger}$
    COLM 2025 [arxiv]

  • Reinforcement Learning with Rubric Anchors
    Zenan Huang*, Yihong Zhuang*, Guoshan Lu*, Zeyu Qin*, Haokai Xu*, Tianyu Zhao, Ru Peng, Jiaqi Hu, Zhanming Shen, Xiaomeng Hu, Xijun Gu, Peiyi Tu, Jiaxin Liu, Wenyu Chen, Yuzhuo Fu, Zhiting Fan, Yanmei Gu, Yuanyuan Wang, Zhengkai Yang, Jianguo Li, Junbo Zhao$^{\dagger}$
    Arxiv 2025 [arxiv]

  • Safety Reasoning with Guidelines
    Haoyu Wang*, Zeyu Qin* $^{\dagger}$, Li Shen, Xueqian Wang, Dacheng Tao, Minhao Cheng.
    ICML 2025 [arxiv]

  • Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (Spotlight!)
    Rui Min*, Zeyu Qin* $^{\dagger}$, Nevin Zhang, Li Shen, Minhao Cheng.
    NeurIPS 2024 [arxiv] [code]

AI Safety

Data Safety

  • Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (Spotlight!)
    Rui Min*, Zeyu Qin* $^{\dagger}$, Nevin Zhang, Li Shen, Minhao Cheng.
    NeurIPS 2024 [arxiv] [code]

  • Towards Stable Backdoor Purification through Feature Shift Tuning
    Rui Min*, Zeyu Qin* $^{\dagger}$, Li Shen, Minhao Cheng.
    NeurIPS 2023 [arxiv] [OpenReview] [code]

  • Revisiting Personalized Federated Learning: Robustness Against Backdoor Attacks
    Zeyu Qin, Liuyi Yao, Daoyuan Chen, Yaliang Li, Boling Ding, Minhao Cheng.
    KDD 2023 [arxiv] [code]

Adversarial Machine Learning

  • Boosting the Transferability of Adversarial Attacks with Reverse Adversarial Perturbation
    Zeyu Qin*, Yanbo Fan*, Yi Liu, Li Shen, Yong Zhang, Jue Wang, Baoyuan Wu.
    NeurIPS 2022 [arxiv] [OpenReview] [code]

  • Random Noise Defense Against Query-Based Black-Box Attacks
    Zeyu Qin, Yanbo Fan, Hongyuan Zha, Baoyuan Wu.
    NeurIPS 2021 [arxiv] [OpenReview] [code]

Collaboration with Excellent Researchers

  • Training-Trajectory-Aware Token Selection
    Zhanming Shen, Jiaqi Hu, Zeyu Qin, Hao Chen, Wentao Ye, Zenan Huang, Yihong Zhuang, Guoshan Lu, Junlin Zhou, Junbo Zhao. Arxiv, 2026 [arxiv]

  • Supervised Fine-Tuning Needs to Unlock the Potential of Token Priority
    Zhanming Shen, Zeyu Qin, Jiaqi Hu, Wentao Ye, Hao Chen, Xiaomeng Hu, Haokai Xu, Gang Chen, Yi R. Fung, Haobo Wang. Arxiv, 2026 [arxiv]

  • UltraHorizon: Benchmarking Agent Capabilities in Ultra Long-Horizon Scenarios
    Haotian Luo*, Huaisong Zhang*, Xuelin Zhang*, Haoyu Wang*, Zeyu Qin*, Wenjie Lu*, Guozheng Ma, Haiying He, Yingsha Xie, Qiyang Zhou, Zixuan Hu, Hongze Mi, Yibo Wang, Naiqiang Tan, Hong Chen, Yi R. Fung, Chun Yuan, Li Shen.
    Arxiv, 2025 [arxiv]

  • Merge-of-Thought Distillation
    Zhanming Shen*, Zeyu Qin*, Zenan Huang, Hao Chen, Jiaqi Hu, Yihong Zhuang, Guoshan Lu, Gang Chen, Junbo Zhao.
    Arxiv, 2025 [arxiv]

  • Lifelong Safety Alignment for Language Models
    Haoyu Wang, Zeyu Qin, Yifei Zhao, Chao Du, Min Lin, Xueqian Wang, Tianyu Pang
    NeurIPS 2025 [arxiv]

  • RoMa: A Robust Model Watermarking Scheme for Protecting IP in Diffusion Models
    Yingsha Xie, Rui Min, Zeyu Qin, Fei Ma, Li Shen, Fei Yu, Xiaochun Cao
    NeurIPS 2025

  • Preserving Diversity in Supervised Fine-Tuning of Large Language Models
    Ziniu Li, Congliang Chen, Tian Xu, Zeyu Qin, Jiancong Xiao, Ruoyu Sun, Zhi-Quan Luo.
    ICLR 2025 [arxiv]

  • Imitation Learning from Imperfection: Theoretical Justifications and Algorithms (Spotlight!)
    Ziniu Li*, Tian Xu*, Zeyu Qin, Yang Yu, Zhiquan Luo.
    NeurIPS 2023 [OpenReview]
    (This is excellent work from Ziniu and Tian. I only conducted part of the experiments. I almost have no idea about Imitation Learning πŸ˜‚.)

πŸŽ– Honors and Awards

  • 2024.09, NeurIPS 2024 Spotlight Paper
  • 2023.09, NeurIPS 2023 Spotlight Paper
  • 2021.04, Poster Runner-Up Reward of The First Doctoral and Postdoctoral Academic Forum of Shenzhen Research Institute of Big Data

πŸ’¬ Invited Talks

  • 2022.10, Adversarial Transferability in AI Times Forum.

πŸ’» Internships

  • September 2025 – Now: Intern, Agent Team of Kimi, Beijing, China
  • November 2024 – July 2025: Research Intern, MSRA GenAI, Beijing, China
  • June 2022 – May 2023: Research Intern, Alibaba Damo Academy, Hangzhou, China
  • July 2021 – May 2022: Research Intern, Tencent AI Lab, Shenzhen, China
  • January 2020 – October 2020: Research Intern, Tencent AI Lab, Shenzhen, China

πŸ“§ Contact

zeyu.qin@connect.ust.hk ; zeyu6181136@gmail.com