Jiayu Ding (丁家钰)
About Me
I am a Master’s student at the School of Electronics and Computer Engineering (SECE) at Peking University, where I have the privilege of being advised by Prof. Ge Li.
My research operates at the intersection of Large Language Models (LLMs) and spatial-temporal data. My goal is to leverage the structured knowledge and reasoning of LLMs to unlock a deeper, more semantic understanding of complex 3D scenes and dynamic videos. I am excited to contribute to building more capable and perceptive AI systems.
Recent News
- 2026/05 - 1 paper accepted to ICML 2026!
- 2026/02 - 1 paper accepted to CVPR 2026!
Selected Publications
(*: Equal Contribution, ✉: Corresponding Author)
[1] ExtrinSplat: Decoupling Geometry and Semantics for Open-Vocabulary Understanding in 3D Gaussian Splatting [Paper]
Jiayu Ding, Xinpeng Liu, Zhiyi Pan, Shiqiang Long, Ge Li✉.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
The first to elevate the understanding primitive of open-vocabulary 3D Gaussian Splatting from geometric points to semantic objects, achieving state-of-the-art performance across multiple benchmarks.
[2] 3D Scene Assertion Verification
Jun Lin*, Jiayu Ding*✉, Xiangtian Si, Xitong Cao, Lixin Hong, Zhang Chen, Chenxi Lv, Wenqian Wang
International Conference on Machine Learning (ICML), 2026
Introduces the 3D Scene Assertion Verification task and the first large-scale 3DSAV benchmark, and proposes the DualLPSS framework with dual-stage routing, achieving state-of-the-art performance on verifying complex logical assertions in 3D scenes.
[3] 3D Instruction Ambiguity Detection [Paper]
Jiayu Ding, Haoran Tang, Hongbo Jin, Wei Gao, Ge Li✉.
Under Review 2026
The first to introduce the task of open-vocabulary instruction ambiguity detection in 3D scene understanding, establishing state-of-the-art results via a novel VLM-based 3D reasoning framework.
[4] VISTA: Mitigating Semantic Inertia in Video-LLMs via Training-Free Dynamic Chain-of-Thought Routing [Paper]
Hongbo Jin*, Jiayu Ding*, Siyi Xie*, Guibo Luo, Ge Li✉.
Under Review 2026
The first to identify “Semantic Inertia” in Video-LLMs where visual evidence is suppressed. By proposing VISTA (a training-free dynamic Chain-of-Thought routing framework), it effectively aligns perception with logic, surpassing base models and rivaling larger proprietary models.
[5] TIR-Flow: Active Video Search and Reasoning with Frozen VLMs [Paper]
Hongbo Jin*, Siyi Xie*, Jiayu Ding*, Kuanwei Lin, Ge Li.
Under Review 2026
Proposes TIR-Flow, a training-free framework that enables frozen VLMs to perform “System-2” active visual search and reasoning, achieving significant gains (+10.5% on Egoschema) without any parameter updates.
Collaboration
I closely collaborate with talented researchers from different universities. Most of these collaborators have contributed actively to our joint projects and achieved tangible research outcomes.
- Jun Lin, China University of Geosciences (2027 Fall, ICML*1)
- Bangpu Chen, China University of Geosciences (2027 Fall, Submitted to ACM MM)
- Xiangtian Si, China University of Geosciences (2029 Fall, Submitted to ECCV)
- Meilu Song, North China Electric Power University (2027 Fall, Submitted to ECCV and ACM MM)
- ShengYao Zhou, Zhejiang University (2027 Fall, Submitted to NeurIPS and EMNLP)
- Hongbo Jin, Peking University (B.S. Huazhong University of Science and Technology)
- Chaoyue Li, Huazhong University of Science and Technology (B.S. China University of Geosciences)
- Zhuodong Liu, Xiamen University (B.S. Beijing Jiaotong University)
- Wenqian Wang, University of Chinese Academy of Sciences (B.S. University of Jinan)
I am always open to new academic exchanges and research collaborations with motivated students and researchers who are interested in 3D vision, video understanding, and visual reasoning. Please feel free to reach out to me at any time if you would like to work together on research projects.
Contact
I maintain an open mind toward cutting-edge technologies.
Research: I warmly welcome students and researchers to reach out for potential collaboration.
Entrepreneurship: I firmly believe that AI Agents will reshape the landscape of education. If you are also exploring the AI4Edu path, I would love to connect and exchange ideas.
📧 Email: jyding25@stu.pku.edu.cn
