About
I am Ganlin Yang, a Ph.D. candidate at the University of Science and Technology of China and a joint PHD student at Shanghai AI Laboratory. My primary research interest is embodied intelligence, especially at the intersection of embodied manipulation and embodied brain models. I aim to develop agents that can perceive, reason, and act coherently in long-horizon tasks, with strong generalization across environments, tasks, and embodiments.
My recent work focuses on how multimodal foundation models can support embodied decision-making through unified end-to-end frameworks. Representative projects include VLASER, Visual Embodied Brain, and EventVLA. These works emphasize spatial intelligence, world-model-guided reasoning, and memory-enhanced policy learning for robust long-horizon behavior.
I also work on multimodal understanding and generation. I have contributed to the InternVL research line, including InternVL3.5, InternVL-U and Intern-S1. This line of work explores model capability scaling, multimodal reasoning, and large-scale open data/model pipelines. Before these directions, I also worked on 3D reconstruction and neural rendering. This experience provides a useful foundation for visual representation learning in embodied systems.
Education
P.H.D candidate of Electronic Engineering & Information Science
School of Information Science and Technology
Supervisor: Prof. Jifeng Dai, Prof. Wengang Zhou
Master of Electronic Engineering & Information Science
School of Information Science and Technology
Supervisor: Prof. Dong Liu
Bachelor of Electronic Engineering & Information Science
School of Gifted Young
GPA: 3.89/4.3 (ranked 10% at School of Gifted Young)
Internship
Research Intern, Intelligent Multimedia Group
Research Topic: 3D reconstruction and Neural Rendering
Supervisor: Dr. Zhizheng Zhang
Research Intern, Multimedia Computing Group
Research Topic: 3D reconstruction and generation
Supervisor: Dr. Jingjing Fu
Research Intern, Large Language Model Center
Research Topic: Multimodal Large Language Model
Supervisor: Dr. Jifeng Dai, Dr. Wenhai Wang
Research Intern, Physical Intelligence Center
Research Topic: Embodied AI; Vision Language Action Model
Supervisor: Dr. Jiangmiao Pang, Dr. Tai Wang
Visiting Intern, ScaleLab
Research Topic: Embodied AI; World-action Model
Supervisor: Dr. Yao Mu
Research Experiences
Embodied Brain and Manipulation
VLASER: Vision-Language-Action Model with Synergistic Embodied Reasoning
ICLR 2026 | [Paper] [GitHub] [Project Page]
Ganlin Yang*, Tianyi Zhang*, Haoran Hao*, Weiyun Wang, Yibin Liu, ..., Wenhai Wang, Yao Mu, Zhi Hou
Summary: VLASER introduces synergistic embodied reasoning to tightly couple scene understanding, instruction grounding, and action prediction for robust long-horizon manipulation.
Visual Embodied Brain: Let Multimodal Large Language Models See, Think, and Control in Spaces
Technical Report | [Paper] [GitHub]
Gen Luo*, Ganlin Yang*, Ziyang Gong*, Guanzhou Chen*, ..., Yu Qiao, Rongrong Ji, Xizhou Zhu
Summary: This work proposes a visual embodied brain paradigm that unifies perception, spatial reasoning, and control planning to improve generality across embodied tasks and environments.
ACE-Brain-0: Spatial Intelligence as a Shared Scaffold for Universal Embodiments
Technical Report | [Paper] [GitHub] [Project Page]
Ziyang Gong, Zehang Luo, Anke Tang, Zhe Liu, Shi Fu, Zhi Hou, Ganlin Yang, ..., Hengshuang Zhao, Dacheng Tao, Xiaogang Wang
Summary: ACE-Brain-0 argues for spatial intelligence as a common abstraction across embodiments, enabling transfer of planning and control priors between heterogeneous robots.
EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies
Technical Report | [Paper] [GitHub] [Project Page]
Ganlin Yang*, Zhangzheng Tu*, Yuqiang Yang*, Sitong Mao, Junyi Dong, Tianxing Chen, Jiaqi Peng, Jing Xiong, Jiafei Cao, Jifeng Dai, Wengang Zhou, Yao Mu, Tai Wang.
Summary: EventVLA introduces event-driven memory updates to preserve key visual evidence during long-horizon interaction, improving temporal consistency and policy robustness.
Cortex: A Bidirectionally Aligned Embodied Agent Framework for Long-horizon Manipulation
Technical Report | [Paper] [GitHub] [Project Page]
Jiaqi Peng*, Xiqian Yu*, Delin Feng*, ..., Ganlin Yang, ..., Jiangmiao Pang, Yuan Shen, Tai Wang.
Summary: Cortex is a bidirectionally aligned embodied agent framework with a customized planning interface that conveys executable and tractable subtask plans from high-level VLM to low-level VLA.
Multimodal Understanding and Generation
InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
Technical Report | [Paper] [GitHub] [Project Page]
Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, ... , Ganlin Yang, ... , Kai Chen, Yu Qiao, Wenhai Wang, Gen Luo.
Summary: InternVL3.5 systematically improves a unified multimodal model across perception, reasoning, and generation, with better scaling behavior and stronger efficiency-quality trade-offs.
InternVL-U: Democratizing Unified Multimodal Models for Understanding, Reasoning, Generation and Editing
Technical Report | [Paper] [GitHub]
Changyao Tian, Danni Yang, Guanzhou Chen, ... , Ganlin Yang, ... , Yu Qiao, Kai Chen, Hongjie Zhang.
Summary: InternVL-U presents a unified multimodal framework that supports both discriminative and generative tasks in one system, enabling broad capability transfer across modalities.
ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework
Technical Report | [Paper] [GitHub]
Guanzhou Chen, Erfei Cui, Changyao Tian, Danni Yang, Ganlin Yang, Yu Qiao, Hongsheng Li, Gen Luo, Hongjie Zhang.
Summary: ScaleEdit-12M builds a multi-agent data engine to generate large-scale editing instruction data, improving data diversity and controllability for open-source image editing models.
3D Reconstruction & Rendering
Drim-NeRF: Diffusion-Based Restoration for Improving Neural Radiance Fields
TCSVT 2025 | [Paper]
Ganlin Yang, Kaidong Zhang, Jingjing Fu, Dong Liu.
Summary: Drim-NeRF introduces a diffusion-based restoration stage to refine degraded views and improve NeRF reconstruction quality under noisy, low-light, or sparsely sampled conditions.
Mask-Based Modeling for Neural Radiance Fields
Ganlin Yang, Guoqiang Wei, Zhizheng Zhang, Yan Lu, Dong Liu.
Summary: This work proposes mask-guided modeling for NeRF training, improving geometry and appearance learning by focusing optimization on informative regions and reducing background-induced artifacts.
Skills
Programming:
Python, MATLAB, C/C++, PyTorch, LaTeX, Linux, Git, Deepspeed, Distributed Training, High-speed Model Inference
Research Skills:
3D Reconstruction and Perception, Multimodal Large Models Understanding and Generation, Reinforcement Learning for Embodied Control, Vision-Language-Action (VLA), World Action Model (WAM), Real-robot Deployment and Evaluation, End-to-end Embodied AI System Integration