Changbin Zhang 张长彬 Ph.D. Candidate The University of Hong Kong zhangchbin@gmail.com Google Scholar

I'm currently a final-year Ph.D. candidate working with Dr. Yujie Zhong and Prof. Kai Han at The University of Hong Kong (HKU) . Before that, I spent wonderful years at Nankai University (NKU), supervised by Prof. Ming-Ming Cheng and received my M.Eng degree in Computer Science.

I am currently working on research related to RL / Agentic RL, on-policy distillation (OPD), Multi-Agent Systems, and self-evolving agents. I have published 5 first-author and 2 co-first-author papers in CCF-A venues, including 1 ESI Highly Cited Paper (Top 1%) and 1 CVPR Highlight paper (Top 2.8%). My work has received over 2,300 citations on Google Scholar. I also won a 🥈 Silver Medal in the ACM-ICPC Asia Regional Contest. I am currently on the job market, seeking roles in post-training and agents.

我目前从事 RL / Agentic RL、on-policy distillation (OPD)、Multi-Agent Systems、self-evolving agents 等相关研究。我以第一作者（含共同一作）身份在 CCF-A 类会议上发表论文 7 篇（5 篇一作、2 篇共同一作），其中包含 1 篇 ESI 高被引论文（Top 1%）和 1 篇 CVPR Highlight 论文（Top 2.8%）。我在 Google Scholar 上累计获得超过 2300 次引用。我也曾获得 ACM-ICPC 亚洲区域赛 🥈 银牌。我目前正在求职，寻找 post-training 与 agent 相关的职位。

Publications

• ICML 2026

  iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning
  Changbin Zhang, Yujie Zhong, Qiang Zhang and Kai Han
  ICML, 2026
  [arXiv] [code] [project page] [models]
  A new third paradigm of visual reasoning that internalizes visual grounding into textual CoT via RL — beyond OpenAI's "think with images" and DeepSeek's "think with visual primitives".

  Reinforcement Learning for MLLMs Think with visual primitives
• Under Review

  Mr. DETR++: Instructive Multi-Route Training for Detection Transformers with Mixture-of-Experts
  Changbin Zhang, Yujie Zhong and Kai Han
  Under Review, 2025
  [arXiv] [code] [project page] [HuggingFace Demo]
  A flexible multi-route MoE training framework that boosts detection transformers to SOTA across object detection, instance and panoptic segmentation.

  MoE Foundation Model
• CVPR 2025

  Mr. DETR: Instructive Multi-Route Training for Detection Transformers
  Changbin Zhang, Yujie Zhong and Kai Han
  CVPR, 2025
  [arXiv] [code] [project page] [HuggingFace Demo]
  April, 2025, Rank #1 in the Leaderboard of COCO 2017 val

  Visual Perception Foundation Model
• CVPR 2025

  v-CLR: View-Consistent Learning for Open-World Instance Segmentation
  Changbin Zhang, Jinhong Ni, Yujie Zhong and Kai Han
  CVPR, 2025, Highlight (Top 2.8% of submissions)
  [arXiv] [code] [project page] [HuggingFace Demo]
  View-consistent learning achieves SOTA on open-world instance segmentation.

  Visual Perception Foundation Model
• CVPR 2023

  Endpoints Weight Fusion for Class Incremental Semantic Segmentation
  Jia-Wen Xiao*, Changbin Zhang*, Jiekang Feng, Xialei Liu, Joost van de Weijer and Ming-Ming Cheng
  CVPR, 2023
  [IEEE/CVF] [code]
  An elegant method that balances stability and plasticity for class-incremental segmentation.

  Visual Perception Foundation Model
• CVPR 2022

  Representation Compensation Networks for Continual Semantic Segmentation
  Changbin Zhang*, Jia-Wen Xiao*, Xialei Liu, Yingcong Chen and Ming-Ming Cheng
  CVPR, 2022
  [IEEE/CVF] [arXiv] [中译版] [code]
  A re-parameterization approach that decouples old/new knowledge, hitting SOTA continual segmentation with zero inference overhead.

  Visual Perception Foundation Model
• IEEE TIP 2021

  Delving Deep into Label Smoothing
  Changbin Zhang*, Peng-Tao Jiang*, Qibin Hou, Yunchao Wei, Qi Han, Zhen Li and Ming-Ming Cheng
  IEEE Transactions on Image Processing (IEEE TIP), 2021
  [IEEE Explore] [arXiv] [中译版] [code]
  An online label smoothing that mines the model's own prediction statistics to craft more reliable soft labels.

  Visual Perception Foundation Model
• IEEE TIP 2021

  LayerCAM: Exploring Hierarchical Class Activation Maps For Localization
  Peng-Tao Jiang*, Changbin Zhang*, Qibin Hou, Ming-Ming Cheng and Yunchao Wei
  IEEE Transactions on Image Processing (IEEE TIP), 2021
  [IEEE Explore] [中译版] [code] [ESI Highly Cited Paper (1%)]
  The first method to extract class activation maps from ANY network layer, revealing hierarchical localization.

  Visual Perception Foundation Model
• ICCV 2021

  Personalized Image Semantic Segmentation
  Yu Zhang, Changbin Zhang, Peng-Tao Jiang, Ming-Ming Cheng and Mao Feng
  ICCV, 2021
  [IEEE/CVF] [arXiv] [中译版] [code]

  Visual Perception Foundation Model
• IEEE TPAMI 2021

  Deep Hough Transform for Semantic Line Detection
  Kai Zhao*, Qi Han*, Changbin Zhang, Jun Xu, and Ming-Ming Cheng
  IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI), 2021
  [IEEE Explore] [arXiv] [中译版] [code] [ESI Highly Cited Paper (1%)]

  Visual Perception Foundation Model

Experience

• 
  ByteDance  ·  Internship  ·  2026.05 ~ present  ·  Beijing
• 
  SAIC Motor Autonomous  ·  Engineer  ·  2022.07 ~ 2023.01  ·  Shanghai, working closely with Dr. Peixuan Li
• 
  NIO Autonomous  ·  Internship  ·  2022.01 ~ 2022.04  ·  Beijing
• 
  DJI  ·  Internship  ·  2021.06 ~ 2021.08  ·  Shenzhen

Honors & Awards

• 💰
  National Scholarship  ·  2017 / 2018  ·  16,000 CNY

  国家奖学金
• 💰
  Nankai - SK Hynix Scholarship  ·  2020 / 2021  ·  42,000 CNY

  南开大学-SK海力士 奖学金
• 🥈
  Silver Medal of International Collegiate Programming Contest (ACM-ICPC) Asia Regional Contest  ·  2017

  ACM-ICPC 亚洲区域赛 银牌
• 🥇
  First Prize of JiangSu Province Collegiate Programming Contest (JSCPC)  ·  2017

  江苏省大学生程序设计竞赛 一等奖