Qisong Zhang (张启淞) — CV/AIGC Researcher CV / AIGC 研究员

📍 Beijing, China 北京  •  🎓 BUPT × China Telecom TeleAI 北邮 × 中国电信  •  ✉️ 2362589239@bupt.edu.cn

🌐 Homepage 学术主页  •  📄 CV 简历  •  ⚡ DataEvolver  •  📄 arXiv:2605.01789  •  🌐 Project Page

"Building goal-driven loop agents & self-evolving data pipelines for multimodal foundation models. — 构建目标驱动的闭环智能体与自演化数据管线，服务多模态基础模型。"

• 2026.08 — Traj-MCPO accepted to ACM MM 2026.（被 ACM MM 2026 录用。）
• 2026.08 — Panel2Page submitted to AAAI 2027 (under review); GeoEdit++ submitted to IEEE TMM (under review).（Panel2Page 提交 AAAI 2027 在审；GeoEdit++ 提交 IEEE TMM 在审。）
• 2026.05 — DataEvolver selected for an Oral Presentation at AIDataSci 2026 (KDD 2026 Workshop) and released on arXiv:2605.01789.（获 AIDataSci 2026 Oral 报告并发布 arXiv。）
• 2026.05 — DataEvolver selected for Poster Presentation at the 2026 BAAI Conference Agent for Science.（入选 2026 BAAI 大会 Agent for Science 海报展示。）
• 2026.03 — Joined China Telecom TeleAI as a data algorithm intern, leading DataEvolver R&D.（加入中国电信 TeleAI 数据算法团队，主导 DataEvolver 研发。）

• Role 角色: First author (equal contribution) · Core R&D — 第一作者（共同一作）· 核心研发
• Core Idea 核心思想: DataEvolver turns controllable visual data construction into recorded feedback loops — candidate generation with VLM/CV quality gates (inner loop), downstream weak-model behavior into targeted training data additions (outer loop).（将可控视觉数据构建转化为可记录的反馈闭环：候选生成与 VLM/CV 质检门控构成内循环，下游弱模型行为转化为定向训练数据补充构成外循环。）
• Key Results 关键成果：
  • 🎤 Oral Presentation 报告 — AIDataSci 2026 (KDD 2026 Workshop)
  • 📜 Poster Presentation 海报展示 — 2026 BAAI Conference Agent for Science Competition
  • 📊 Downstream Impact 下游效果: Reached 16.79 PSNR, 0.9525 CLIP-I, and 0.8918 DINO on 488 image-level rotation samples of SpatialEdit, outperforming Public LoRA by +1.03 dB PSNR.（在 SpatialEdit 的 488 个图像级物体旋转样本上达到 16.79 PSNR、0.9525 CLIP-I、0.8918 DINO，较 Public LoRA 提升 +1.03 dB PSNR。）

• 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
  AIDataSci 2026 · KDD 2026 Workshop (Oral) · arXiv:2605.01789 — First author (equal contribution) 第一作者（共同一作）, * equal contribution 共同一作
  [ arXiv ] · [ Code ] · [ Project Page ] · [ OpenReview ] · [ Method 方法图 ]

• 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
  ACM MM 2026 (Accepted 已录用) · Second author 第二作者
  [ Code ]
  A trajectory-wise policy optimization framework for stable RL post-training of rectified-flow text-to-image models.（面向 Rectified-Flow 文生图模型的轨迹级边际一致性策略优化。）

• GeoEdit++: Learning Residual-Conditioned Rotation References for Geometric Image Editing
  Qisong Zhang et al.
  IEEE Transactions on Multimedia (Under Review 在审) · Co-author 共同作者
  Residual-conditioned rotation references for geometric image editing.（残差条件旋转参考学习，提升复杂背景下 3D 物体旋转编辑的真实感与几何对齐。）

• DataEvolver — Goal-Driven Synthetic Visual Data Engine 目标驱动的闭环视觉数据自演化引擎
  VLM + Agent rendering loop with a 24-action space, inner quality gates, and an outer downstream feedback loop that turns weak angles into targeted data updates.（24 有界动作空间、内层质量门控与下游反馈外循环，将弱视角转化为定向数据扩增。）（method 方法图 · evolution 自演化 · comparison 对比）
• Multi-task Image Editing LoRA System 多任务图像编辑 LoRA 微调系统
  Fine-tuned 30B Qwen-Image-Edit for rotation, movement, insertion, and removal; trained 11 angle-specific LoRAs (30°–180°) with a Preset-LoRA curriculum & additive rehearsal against forgetting.（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，Preset LoRA 课程学习 + Additive Rehearsal 抗遗忘。）
• YouAI — Self-Evolving Multimodal UI Agent 自演化多模态 UI 智能体
  Manager/Operator architecture based on Mobile-Agent-E with persistent Tips/Shortcuts memory; 86.9% satisfaction & 90.4% action accuracy on Mobile-Eval-E.（基于 Mobile-Agent-E 的 Manager/Operator 架构，加入 Tips/Shortcuts 长期经验记忆；Mobile-Eval-E 上 86.9% 满意度、90.4% 动作准确率。）
• Graduate Education Assessment Assistant 研究生教育质量评估助手
  LLM-assisted subjective scoring (correctness, logic, expression) with human-in-the-loop review & ChromaDB RAG.（LLM 辅助主客观协同评分 + ChromaDB RAG 检索。）

• 2026 — Oral Presentation at AIDataSci 2026 (KDD 2026 Workshop): DataEvolver.（Oral 报告展示。）
• 2026 — Poster Presentation at the 2026 BAAI Conference Agent for Science Competition: DataEvolver.（海报展示。）