Cognitive and Physical Intelligence
How can machines model, learn, and continually evolve cognitive and physical capabilities in the human world, both digital and physical?
AI agents
Wenwei Zhang 张文蔚
Sr. Staff Research Scientist · XPENG Robotics
My research focuses on scalable oversight of agents in the physical world at XPENG Robotics.
Previously, at Shanghai AI Laboratory, I was responsible for the post-training of the InternLM and Intern-S model series and conducted research on scalable oversight of digital agents. As a core member of OpenMMLab since 2019, I built and released MMDetection3D and have led the development of MMDetection and MMDetection3D since 2020. I received my Ph.D. from Nanyang Technological University, Singapore, where I worked at MMLab under the supervision of Professor Chen Change (Cavan) Loy. I have also collaborated with Kai Chen, Jiangmiao Pang, and Jianping Shi.
Research Direction
Scalable Oversight
I am interested in how agents develop cognitive and physical intelligence and how people can remain valuable collaborators as agents evolve. Reinforcement learning offers a key mechanism for this evolution through a closed loop of interaction, feedback, and policy improvement. Specifically, I study the following two questions:
Human-Machine Collaboration
How can evolving machine cognition assist complex human activities while keeping humans valuable in an increasingly automated world?
I approach these two questions through scalable oversight, providing reliable feedback at scale for reinforcement learning while keeping people involved in guiding agents’ progress and working alongside them. I view self-improvement and recursive self-improvement (RSI) as natural outcomes of making this learning loop reliable and scalable.
Open source
Selected projects
Foundation Models
Foundation model-based Agents
Long-horizon Reasoning Agent
Intern-S2 supported a human–AI collaboration that proved the three-dimensional Ghys conjecture.
Ghys · People’s Daily OnlineComputer Vision Algorithm System
Recent News
Selected updates
-
Joined XPENG Robotics as a Sr. Staff Research Scientist and Sr. Manager, leading the Physical Agents team.
-
Intern-S2 supported a human–AI collaboration that proved the three-dimensional Ghys conjecture. Coverage in People’s Daily Online.
-
Appointed Head of the Large Model Center at Shanghai AI Laboratory.
-
We released Intern-S1-Pro, an open-source trillion-parameter multimodal foundation model for general and scientific reasoning. Technical report.
-
Intern-S1-MO achieved gold-medal-level performance, ranking 3rd among human competitors and 1st among AI competitors in the Chinese Olympiad in Mathematics (CMO2025).
-
Intern-S1, state-of-the-art open-source multi-modal foundation model has been released. See Intern-S1 Collections and tech report.
Earlier updates 2025–2019
-
InternLM2.5-StepProver received a Best Paper Honorable Mention at the ICML 2025 AI4Math Workshop.
-
InternLM2.5-Chat, state-of-the-art open-source LLMs in opencompass and huggingface leadboard, have been released. See InternLM2.5 Model Zoo.
-
InternLM2-Chat, state-of-the-art open-source LLMs in opencompass and AlpacaEval2, have been released. See InternLM2 Model Zoo.
-
2 papers (CLIPSelf and UniHSI) are accepted as spotlights by ICLR2024.
-
We release OpenMMLab 2.0 with a new core, MMEngine.
-
Video K-Net is accepted by CVPR 2022 (oral).
-
K-Net is accepted by NeurIPS 2021. Code has been released at here.
-
MMDet3D team obtains Best PKL Award and best vision-only results in the 3rd nuScenes detection challenge of 5th AI Driving Olympics, NeurIPS 2020. Paper of our multi-modality method is released in arxiv.
-
Second runner up in LVIS2020 Challenge. Paper can be found in arxiv.
-
We release MMDetection3D, OpenMMLab's next-generation platform for general 3D object detection.
-
Win the 1st prize in COCO 2019 Object Detection Challenge (no external data). (Team: MMDet)
Academic Service