Qisong Zhang · 张启淞
CV / AIGC Researcher · Beijing University of Posts and Telecommunications & China Telecom TeleAI
I am an incoming M.S. student at Beijing University of Posts and Telecommunications (BUPT), jointly trained with China Telecom's Institute of Artificial Intelligence (TeleAI) under the Excellent Engineer Program. My research focuses on On-Policy Distillation (OPD) for image and video generation models, controllable world generation, and goal-driven synthetic data engines.
我是北京邮电大学与中国电信人工智能研究院 (TeleAI) 卓越工程师联合培养计划的待入学硕士研究生。研究方向聚焦于图像与视频生成模型的 On-Policy 蒸馏 (OPD)、可控视觉世界模型，以及目标驱动的合成数据闭环演化引擎。
Feel free to reach out via email or GitHub if you are interested in collaboration or potential opportunities.
Research Interests
On-Policy Distillation (OPD) World Generation & Narrative Synthetic Data & Geometric Editing Multimodal UI & Action Agents
News
- 2026.08 Traj-MCPO Accepted to ACM MM 2026
Traj-MCPO (trajectory-wise marginal consistent policy optimization) was accepted to ACM MM 2026. 2026年8月 Traj-MCPO 被 ACM MM 2026 录用
Traj-MCPO（轨迹级边际一致性策略优化）论文被 ACM MM 2026 正式录用。 - 2026.08 Panel2Page Submitted to AAAI 2027 & GeoEdit++ to IEEE TMM
New revision of Panel2Page (visual narrative generation) posted to AAAI 2027; GeoEdit++ (residual rotation references) submitted to IEEE TMM. 2026年8月 Panel2Page 提交 AAAI 2027、GeoEdit++ 提交 IEEE TMM 期刊在审
Panel2Page 漫画单页视觉叙事生成论文完成 AAAI 2027 审稿更新，GeoEdit++ 几何图像编辑论文提交 IEEE TMM 期刊在审。 - 2026.05 DataEvolver KDD Workshop Oral & arXiv Release
Paper selected for an Oral Presentation at AIDataSci 2026 (KDD 2026 Workshop) and released on arXiv:2605.01789. 2026年5月 DataEvolver 获 AIDataSci (KDD Workshop) Oral 报告并发布 arXiv
论文被 KDD 2026 Workshop (AIDataSci 2026) 选为 Oral 报告展示，并在 arXiv:2605.01789 正式发布。 - 2026.05 BAAI Agent for Science Poster Presentation
DataEvolver was selected for on-site Poster Presentation in the 2026 BAAI Conference Agent for Science Competition. 2026年5月 DataEvolver 入选 2026 BAAI 大会 Agent for Science 海报展示
DataEvolver 入选 2026 北京智源大会 Agent for Science 竞赛现场海报展示并获颁证书。 - 2026.03 Joined TeleAI Data Algorithm Team
Joined Institute of Artificial Intelligence (TeleAI) as a data algorithm intern leading DataEvolver. 2026年3月 加入 TeleAI 数据算法团队开展 DataEvolver 研发
加入中国电信人工智能研究院 (TeleAI) 担任数据算法实习生，主导 DataEvolver 核心研发。
Experience
Beijing University of Posts and Telecommunications
2026.09 - 2029.07
Incoming M.S. Student · Excellent Engineer Program (Jointly with TeleAI)
2026.09 - 2029.07
Incoming M.S. Student · Excellent Engineer Program (Jointly with TeleAI)
北京邮电大学 (BUPT)
2026.09 - 2029.07
待入学硕士研究生 · 卓越工程师联合培养计划（与中国电信 TeleAI 联合培养）
2026.09 - 2029.07
待入学硕士研究生 · 卓越工程师联合培养计划（与中国电信 TeleAI 联合培养）
Institute of Artificial Intelligence (TeleAI), China Telecom
2026.03 - 2026.05
Data Algorithm Intern · DataEvolver Lead Developer
2026.03 - 2026.05
Data Algorithm Intern · DataEvolver Lead Developer
中国电信人工智能研究院 (TeleAI)
2026.03 - 2026.05
数据算法实习生 · DataEvolver 项目核心研发
2026.03 - 2026.05
数据算法实习生 · DataEvolver 项目核心研发
Institute of Artificial Intelligence (TeleAI), China Telecom
2025.09 - Present
AI Algorithm Intern · Multi-task Image Editing LoRA Team
2025.09 - Present
AI Algorithm Intern · Multi-task Image Editing LoRA Team
中国电信人工智能研究院 (TeleAI)
2025.09 - 至今
AI 算法实习生 · 多任务图像编辑 LoRA 项目组
2025.09 - 至今
AI 算法实习生 · 多任务图像编辑 LoRA 项目组
Beijing University of Posts and Telecommunications
2022.09 - 2026.07
B.Eng. Candidate in Information Engineering · School of Artificial Intelligence
2022.09 - 2026.07
B.Eng. Candidate in Information Engineering · School of Artificial Intelligence
北京邮电大学 (BUPT)
2022.09 - 2026.07
信息工程工学学士 · 人工智能学院
2022.09 - 2026.07
信息工程工学学士 · 人工智能学院
Publications
(* equal contribution · † corresponding author · ‡ project leader)
DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents
Qisong Zhang*, Wenzhuo Wu*, Zhuangzhuang Jia, Yunhao Yang, Shuo Zhang, Huayu Zhang, Xianghao Zang, Zhixiang He, Zhongjiang He, Kongming Liang, Zhanyu Ma.
First author (equal contribution) · Core R&D
Designed a closed-loop visual data engine with VLM/CV review, 24 bounded actions, and inner/outer feedback loops. Ours+DualGate improved SpatialEdit PSNR by +1.03 dB over Public LoRA.
AIDataSci 2026, KDD 2026 Workshop (Oral) · arXiv:2605.01789 Oral [arXiv] [code] [project page] [OpenReview] [method]
Qisong Zhang*, Wenzhuo Wu*, Zhuangzhuang Jia, Yunhao Yang, Shuo Zhang, Huayu Zhang, Xianghao Zang, Zhixiang He, Zhongjiang He, Kongming Liang, Zhanyu Ma.
First author (equal contribution) · Core R&D
Designed a closed-loop visual data engine with VLM/CV review, 24 bounded actions, and inner/outer feedback loops. Ours+DualGate improved SpatialEdit PSNR by +1.03 dB over Public LoRA.
AIDataSci 2026, KDD 2026 Workshop (Oral) · arXiv:2605.01789 Oral [arXiv] [code] [project page] [OpenReview] [method]
DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents
张启淞*, Wenzhuo Wu*, Zhuangzhuang Jia, Yunhao Yang, Shuo Zhang, Huayu Zhang, Xianghao Zang, Zhixiang He, Zhongjiang He, Kongming Liang, Zhanyu Ma.
第一作者（共同一作）· 核心研发
设计包含 VLM/CV 评审门控与 24 有界动作空间的闭环数据演化引擎，通过内外双循环将下游客观指标 PSNR 相比 Public LoRA 提升 1.03 dB。
AIDataSci 2026, KDD 2026 Workshop (Oral 报告) · arXiv:2605.01789 Oral [arXiv] [code] [项目主页] [OpenReview] [方法图]
张启淞*, Wenzhuo Wu*, Zhuangzhuang Jia, Yunhao Yang, Shuo Zhang, Huayu Zhang, Xianghao Zang, Zhixiang He, Zhongjiang He, Kongming Liang, Zhanyu Ma.
第一作者（共同一作）· 核心研发
设计包含 VLM/CV 评审门控与 24 有界动作空间的闭环数据演化引擎，通过内外双循环将下游客观指标 PSNR 相比 Public LoRA 提升 1.03 dB。
AIDataSci 2026, KDD 2026 Workshop (Oral 报告) · arXiv:2605.01789 Oral [arXiv] [code] [项目主页] [OpenReview] [方法图]
Panel2Page: Structured Visual Narrative Generation with Omni-Panel Alignment
Shuo Zhang, Qisong Zhang, LanXiang Zhou, Huayu Zhang, Xianghao Zang, Zhixiang He, Kongming Liang, Zhanyu Ma.
Second author
Proposes Panel2Page, a two-stage framework for text-to-comic page generation that separates global composition learning (Page-Level SFT) from hierarchical semantic alignment (Omni-Panel Alignment, OPA), trained on ~20,000 verified prompt-page pairs.
AAAI 2027 (Under Review) Under Review [OpenReview]
Shuo Zhang, Qisong Zhang, LanXiang Zhou, Huayu Zhang, Xianghao Zang, Zhixiang He, Kongming Liang, Zhanyu Ma.
Second author
Proposes Panel2Page, a two-stage framework for text-to-comic page generation that separates global composition learning (Page-Level SFT) from hierarchical semantic alignment (Omni-Panel Alignment, OPA), trained on ~20,000 verified prompt-page pairs.
AAAI 2027 (Under Review) Under Review [OpenReview]
Panel2Page: Structured Visual Narrative Generation with Omni-Panel Alignment
Shuo Zhang, 张启淞, LanXiang Zhou, Huayu Zhang, Xianghao Zang, Zhixiang He, Kongming Liang, Zhanyu Ma.
第二作者
提出 Panel2Page 两阶段框架用于结构化视觉叙事（漫画单页）生成，通过页面级 SFT 学习全局布局先验，并引入 Omni-Panel Alignment (OPA) 实现无额外推理开销的多格层次化语义对齐。
AAAI 2027 在审提交 在审 [OpenReview]
Shuo Zhang, 张启淞, LanXiang Zhou, Huayu Zhang, Xianghao Zang, Zhixiang He, Kongming Liang, Zhanyu Ma.
第二作者
提出 Panel2Page 两阶段框架用于结构化视觉叙事（漫画单页）生成，通过页面级 SFT 学习全局布局先验，并引入 Omni-Panel Alignment (OPA) 实现无额外推理开销的多格层次化语义对齐。
AAAI 2027 在审提交 在审 [OpenReview]
GeoEdit++: Learning Residual-Conditioned Rotation References for Geometric Image Editing
Qisong Zhang et al.
Co-author
Proposes residual-conditioned rotation reference learning for geometric image editing, improving 3D object rotation fidelity and structural alignment across complex backgrounds.
IEEE Transactions on Multimedia (Under Review) Under Review
Qisong Zhang et al.
Co-author
Proposes residual-conditioned rotation reference learning for geometric image editing, improving 3D object rotation fidelity and structural alignment across complex backgrounds.
IEEE Transactions on Multimedia (Under Review) Under Review
GeoEdit++: Learning Residual-Conditioned Rotation References for Geometric Image Editing
张启淞 et al.
共同作者
提出用于几何图像编辑的残差条件旋转参考学习机制，有效提升复杂背景下 3D 物体旋转编辑的真实感与几何对齐精度。
IEEE Transactions on Multimedia (TMM 期刊在审) 在审
张启淞 et al.
共同作者
提出用于几何图像编辑的残差条件旋转参考学习机制，有效提升复杂背景下 3D 物体旋转编辑的真实感与几何对齐精度。
IEEE Transactions on Multimedia (TMM 期刊在审) 在审
Traj-MCPO: Trajectory-wise Marginal Consistent Policy Optimization
Shuo Zhang, Qisong Zhang, Huayu Zhang, LanXiang Zhou, Xianghao Zang, Zhongjiang He, Zhixiang He, Chi Zhang, Tianwei Cao, Kongming Liang, Zhanyu Ma.
Second author
Proposes trajectory-wise marginal consistent policy optimization for rectified-flow T2I alignment, using length-normalized trajectory likelihood ratios and clipping to resolve training-inference marginal mismatch.
ACM MM 2026 (Accepted) Accepted
Shuo Zhang, Qisong Zhang, Huayu Zhang, LanXiang Zhou, Xianghao Zang, Zhongjiang He, Zhixiang He, Chi Zhang, Tianwei Cao, Kongming Liang, Zhanyu Ma.
Second author
Proposes trajectory-wise marginal consistent policy optimization for rectified-flow T2I alignment, using length-normalized trajectory likelihood ratios and clipping to resolve training-inference marginal mismatch.
ACM MM 2026 (Accepted) Accepted
Traj-MCPO: Trajectory-wise Marginal Consistent Policy Optimization
Shuo Zhang, 张启淞, Huayu Zhang, LanXiang Zhou, Xianghao Zang, Zhongjiang He, Zhixiang He, Chi Zhang, Tianwei Cao, Kongming Liang, Zhanyu Ma.
第二作者
针对 Rectified-Flow 文生图模型提出轨迹级边际一致性策略优化方法，利用长度归一化的轨迹似然比与轨迹剪切消除训练与推理之间的边际分布失配。
ACM MM 2026（已录用） 已录用
Shuo Zhang, 张启淞, Huayu Zhang, LanXiang Zhou, Xianghao Zang, Zhongjiang He, Zhixiang He, Chi Zhang, Tianwei Cao, Kongming Liang, Zhanyu Ma.
第二作者
针对 Rectified-Flow 文生图模型提出轨迹级边际一致性策略优化方法，利用长度归一化的轨迹似然比与轨迹剪切消除训练与推理之间的边际分布失配。
ACM MM 2026（已录用） 已录用
- KDD Wksp Oral DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents
Oral Presentation. [arXiv][code][project] KDD Wksp Oral DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents
Oral 报告。 [arXiv][代码][项目主页] - AAAI 2027 Panel2Page: Structured Visual Narrative Generation with Omni-Panel Alignment
Under Review. [OpenReview] AAAI 2027 Panel2Page: Structured Visual Narrative Generation with Omni-Panel Alignment
在审。 [OpenReview] - IEEE TMM GeoEdit++: Learning Residual-Conditioned Rotation References for Geometric Image Editing
Under Review. IEEE TMM GeoEdit++: Learning Residual-Conditioned Rotation References for Geometric Image Editing
在审。 - ACM MM 2026 Traj-MCPO: Trajectory-wise Marginal Consistent Policy Optimization
Accepted. ACM MM 2026 Traj-MCPO: Trajectory-wise Marginal Consistent Policy Optimization
已录用。
Projects
DataEvolver: Goal-Driven Synthetic Visual Data Engine
AIDataSci 2026 Oral / BAAI Poster · 2026
VLM + Agent rendering loop with a 24-action space, inner quality gates, and an outer downstream feedback loop that converts weak angles into targeted data updates. Achieved +1.03 dB PSNR over Public LoRA.
Project [paper] [code] [project page] [method] [evolution] [comparison]
AIDataSci 2026 Oral / BAAI Poster · 2026
VLM + Agent rendering loop with a 24-action space, inner quality gates, and an outer downstream feedback loop that converts weak angles into targeted data updates. Achieved +1.03 dB PSNR over Public LoRA.
Project [paper] [code] [project page] [method] [evolution] [comparison]
Multi-task Image Editing LoRA System
TeleStudio & TeleAI Internship · 2025
Fine-tuned 30B Qwen-Image-Edit for object rotation, movement, insertion, and removal. Trained 11 angle-specific LoRA models (30°-180°), built a Preset LoRA curriculum, and used Additive Rehearsal to prevent forgetting.
Project
TeleStudio & TeleAI Internship · 2025
Fine-tuned 30B Qwen-Image-Edit for object rotation, movement, insertion, and removal. Trained 11 angle-specific LoRA models (30°-180°), built a Preset LoRA curriculum, and used Additive Rehearsal to prevent forgetting.
Project
多任务图像编辑 LoRA 微调系统
TeleStudio & TeleAI 实习成果 · 2025
基于 30B Qwen-Image-Edit 微调旋转、移动、插入与消除任务。训练 11 个视角特定 LoRA 模型 (30°-180°)，构建 Preset LoRA 课程学习与 Additive Rehearsal 机制克服灾难性遗忘。
项目
TeleStudio & TeleAI 实习成果 · 2025
基于 30B Qwen-Image-Edit 微调旋转、移动、插入与消除任务。训练 11 个视角特定 LoRA 模型 (30°-180°)，构建 Preset LoRA 课程学习与 Additive Rehearsal 机制克服灾难性遗忘。
项目
YouAI: Self-Evolving Multimodal UI Agent
Mobile Automation System · 2025
Manager/Operator architecture based on Mobile-Agent-E for long-horizon mobile automation. Added persistent Tips and Shortcuts memory, achieving 86.9% satisfaction and 90.4% action accuracy on Mobile-Eval-E.
Project
Mobile Automation System · 2025
Manager/Operator architecture based on Mobile-Agent-E for long-horizon mobile automation. Added persistent Tips and Shortcuts memory, achieving 86.9% satisfaction and 90.4% action accuracy on Mobile-Eval-E.
Project
YouAI：自演化多模态 UI 智能体
移动端自动化智能体系统 · 2025
基于 Mobile-Agent-E 演进 Manager/Operator 架构，加入 Tips/Shortcuts 长期经验记忆库，在 Mobile-Eval-E 基准上达到 86.9% 满意度与 90.4% 动作准确率。
项目
移动端自动化智能体系统 · 2025
基于 Mobile-Agent-E 演进 Manager/Operator 架构，加入 Tips/Shortcuts 长期经验记忆库，在 Mobile-Eval-E 基准上达到 86.9% 满意度与 90.4% 动作准确率。
项目
Graduate Education Assessment Assistant
RAG & PEFT Assisted Scoring · 2024
LLM-assisted educational scoring system with weighted subjective evaluation (correctness, logic, expression). Combined automatic objective grading with human-in-the-loop review and ChromaDB RAG.
Project
RAG & PEFT Assisted Scoring · 2024
LLM-assisted educational scoring system with weighted subjective evaluation (correctness, logic, expression). Combined automatic objective grading with human-in-the-loop review and ChromaDB RAG.
Project
研究生教育质量主客观协同评估助手
RAG 与 PEFT 辅助评分 · 2024
LLM 辅助教育评估系统，包含正确性、逻辑性与表达清晰度的加权主观打分，结合客观题自动批改、人机协同复核与 ChromaDB RAG 检索。
项目
RAG 与 PEFT 辅助评分 · 2024
LLM 辅助教育评估系统，包含正确性、逻辑性与表达清晰度的加权主观打分，结合客观题自动批改、人机协同复核与 ChromaDB RAG 检索。
项目
Honors & Awards
- 2026 — Oral Presentation, AIDataSci 2026 (KDD 2026 Workshop): DataEvolver. 2026 — AIDataSci 2026 (KDD 2026 Workshop) Oral 报告展示：DataEvolver。
- 2026 — Poster Presentation, 2026 BAAI Conference Agent for Science Competition: DataEvolver. 2026 — 2026 北京智源大会 Agent for Science 竞赛海报展示：DataEvolver。
Poster Presentation Certificate, 2026 BAAI Conference Agent for Science
Talks
- 2026 — “DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents”, Oral Presentation at AIDataSci 2026 (KDD 2026 Workshop). 2026 — 《DataEvolver: Let Your Data Build and Improve Itself via Goal-Driven Loop Agents》，AIDataSci 2026 (KDD 2026 Workshop) Oral 报告。
- 2026 — “DataEvolver”, Poster Presentation, 2026 BAAI Conference Agent for Science Competition. 2026 — 《DataEvolver》，2026 北京智源大会 Agent for Science 竞赛海报展示。