Shusheng Yang

Early research

Yang's earlier papers were on video and object-level vision with transformers. As first or co-first author, Yang worked on video instance segmentation, including "Crossover Learning for Fast Online Video Instance Segmentation" and "Instances as Queries" (ICCV 2021) and the Temporally Efficient Vision Transformer (CVPR 2022 oral), and on masked image modeling, including "Masked Visual Reconstruction in Language Semantic Space" (CVPR 2023) and work on plain vision transformers for object detection (ICCV 2023).¹

Multimodal models and spatial intelligence

Yang was a co-first author of the 2023 Qwen-VL technical report, which introduced the Qwen-VL and Qwen-VL-Chat vision-language models. Since starting a PhD at NYU Courant in September 2024, advised by Saining Xie, Yang has focused on spatial intelligence and long-context video. Yang co-authored the open Cambrian-1 multimodal models (NeurIPS 2024 oral), was co-first author of Thinking in Space (CVPR 2025 oral), which introduced the VSI-Bench benchmark, and was first author of Cambrian-S (ICLR 2026), on spatial supersensing in video.⁵¹⁴³²

Other recent co-authored papers include VideoNSA (ICLR 2026), which applies native sparse attention to video understanding, and a COLM 2026 paper arguing that benchmark designers should train on their own test sets to expose shortcuts that do not require vision.¹

Work and contributions

• Cambrian-S Open-source code

  First author (equal contribution with Jihan Yang) of "Cambrian-S: Towards Spatial Supersensing in Video" (2025; ICLR 2026), which presents the VSI-SUPER benchmark, curates the VSI-590K dataset and trains Cambrian-S models; the code repository is licensed under Apache 2.0.²¹⁶

• Thinking in Space (VSI-Bench) Open-source code

  Co-first author of "Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces" (CVPR 2025 oral), which introduced VSI-Bench, a video-based visual-spatial intelligence benchmark; the code repository is licensed under Apache 2.0.³¹⁷

• Cambrian-1 Open weights

  Co-author of "Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs" (NeurIPS 2024 oral), whose authors provide model weights, code, supporting tools and datasets; the code repository is licensed under Apache 2.0.⁴¹⁸

• Qwen-VL Open weights

  Co-first author with Jinze Bai and Shuai Bai of the Qwen-VL technical report (2023), which introduces the Qwen-VL and Qwen-VL-Chat vision-language models and says code, demo and models are available; the repository's license file is the Tongyi Qianwen license agreement.⁵¹⁹

Timeline

1. 2021Presented two papers on video instance segmentation at ICCV 2021 as co-first author.¹
2. 2022Presented the Temporally Efficient Vision Transformer for video instance segmentation as a CVPR 2022 oral.¹
3. Aug 2023Posted the Qwen-VL technical report as a co-first author.⁵
4. Jun 2024Co-authored Cambrian-1, a fully open, vision-centric family of multimodal language models.⁴
5. Sep 2024Began a PhD at NYU Courant advised by Saining Xie.¹
6. Dec 2024Posted Thinking in Space, introducing the VSI-Bench visual-spatial benchmark, as co-first author.³
7. Nov 2025Released Cambrian-S, on spatial supersensing in video, as first author.²¹
8. Jan 2026Cambrian-S and VideoNSA were accepted to ICLR 2026.¹

Where they work

• NYU 12 people in the wiki

Related people

• Saining XieAlso at NYU
• Yann LeCunAlso at NYU
• Matthew ZeilerAlso at NYU
• Graham TaylorAlso at NYU
• Soumith ChintalaAlso at NYU
• Xiaodong HeMultimodal models · Computer vision
• João CarreiraComputer vision · Multimodal models
• Ali FarhadiComputer vision · Multimodal models
• Aditya RameshMultimodal models · Computer vision
• Xinlei ChenComputer vision · Multimodal models