Zhicheng Wang

I am currently a staff research engineer in Google DeepMind. I am a core contributor of Gemini on vision area, working on Pretraining, SFT and RL. Recently I am focus on video spatial understanding in pre-training and agentic vision tool use / GDPVal in post training. Before LLM/VLM, I mainly worked on 3D computer vision. I graduated from Tsinghua University and UCSD. I had several AI/AR startup experiences.

Email / Scholar / X / LinkedIn

Research & Projects

	The Cartesian Shortcut: Re-evaluate Vision Reasoning in Polar Coordinate Space Xia Hu, Zhenrui Yue, Brian Potetz, Howard Zhou, Leonidas Guibas, Chun-Ta Lu, Zhicheng Wang arXiv, 2026 arXiv introduce Polaris-Bench, a benchmark that dismantles models' reliance on grid-based textual shortcuts by shifting 53 visual reasoning tasks from Cartesian to Polar coordinates, revealing a dramatic performance collapse from up to 83% down to 39% despite identical logical constraints.
	Image Generators are Generalist Vision Learners Valentin Gabeur, Shangbang Long, Songyou Peng, Paul Voigtlaender, Shuyang Sun, Yanan Bao, Karen Truong, Zhicheng Wang, Wenlei Zhou, Jonathan T. Barron, Kyle Genova, Nithish Kannen, Sherry Ben, Yandong Li, Mandy Guo, Suhas Yogin, Yiming Gu, Huizhong Chen, Oliver Wang, Saining Xie, Howard Zhou, Kaiming He, Thomas Funkhouser, Jean-Baptiste Alayrac, Radu Soricut arXiv, 2026 project page / arXiv Reducing monodepth, normals, and semantic segmentation tasks to a single RGB-image-prediction task and fine-tuning an image generator (Nano Banana Pro) beats or ~matches specialized models.
	Gemini 3 2026 project page Best for frontier intelligence at speed
	Robot Learning from a Physical World Model Jiageng Mao, Sicheng He, Hao-Ning Wu, Yang You, Shuyang Sun, Zhicheng Wang, Yanan Bao, Huizhong Chen, Leonidas Guibas, Vitor Guizilini, Howard Zhou, Yue Wang ICRA, 2026 project page / code PhysWorld unifies video generation and real-to-sim world modeling for zero-shot robotic manipulation.
	GaussianLens: Localized High-Resolution Reconstruction via On-Demand Gaussian Densification Yijia Weng, Zhicheng Wang, Songyou Peng, Saining Xie, Howard Zhou, Leonidas Guibas ECCV, 2026 (oral) project page / arXiv GaussianLens uses localized Gaussian densification to reconstruct high-resolution geometry on demand.
	Gemini 2.5 2025 project page / arXiv Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
	ConDense: Consistent 2D/3D Pre-training for Dense and Sparse Features from Multi-View Images Xiaoshuai Zhang, Zhicheng Wang, Howard Zhou, Soham Ghosh, Danushen Gnanapragasam, Varun Jampani, Hao Su, Leonidas Guibas ECCV, 2024 (Oral) arXiv / code Consistent 2D/3D pre-training framework that jointly learns dense and sparse features from multi-view images using 2D-3D feature consistency.
	Towards Degradation-Robust Reconstruction in Generalizable NeRF Chan Ho Park, Ka Leong Cheng, Zhicheng Wang, Qifeng Chen arXiv, 2024 arXiv Degradation-robust generalizable NeRF with depth-aware denoising for handling blurred and degraded input images in 3D reconstruction.
	IBRNet: Learning Multi-View Image-Based Rendering Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, Thomas Funkhouser CVPR, 2021 project page / code / arXiv By learning how to pay attention to input images at render time, we can amortize inference for view synthesis and reduce error rates by 15%.
	MediaPipe KNIFT: Template-based Feature Matching Zhicheng Wang, Genzhi Ye Google Blog, 2020 blog post Template-based feature matching algorithm optimized for mobile and real-time applications.
	Weakly Supervised Glasses Removal Zhicheng Wang, Yisu Zhou, Lijie Wen ICGIP, 2014 PDF Weakly supervised method to remove eyeglasses from face images using sparse coding and optical flow.