Xinyuan Wang

Hi! I am Xinyuan Wang (王心远).

I am a Ph.D. candidate at the University of Hong Kong. I work on agentic foundation models, especially computer-use agent models.

News

All news

Oct 7, 2026
RobotWorld is on arXiv! A new testbed for multimodal agents across robot tasks and embodiments.
Jun 28, 2026
Introducing OSWorld 2.0: 108 long-horizon workflows for evaluating computer-use agents on realistic tasks.
Jun 22, 2026
PhoneBuddy and GameCraft-Bench are on arXiv — exploring how agents use phones and build playable games.
Jan 26, 2026
Computer Agent Arena and VideoAgentTrek are accepted by ICLR 2026!
Oct 11, 2025
🎉 OpenCUA received the Best Paper Award at the COLM AIA Workshop!
Sep 19, 2025
OpenCUA and Jedi are accepted by NeurIPS as Spotlight paper!
May 19, 2025
Scaling Computer-Use Grounding via User Interface Decomposition and Synthesis is published on Arxiv! We introduce OSWorld-G, a large-scale grounding benchmark, and Jedi, a strong grounding model.
Apr 8, 2024
LLM Reasoners: New Evaluation, Library, and Analysis of Step-by-Step Reasoning with Large Language Models is accepted by ICLR2024 workshop!

Selected Publications

All publications

NeurIPS 2025 NeurIPS 2025 Spotlight Best Paper Award, COLM 2025 AIA Workshop

OpenCUA: Open Foundations for Computer-Use Agents

Xinyuan Wang, Bowen Wang, Dunjie Lu, Junlin Yang, Tianbao Xie

arXiv PDF Code Website Poster

Vision-language models have demonstrated impressive capabilities as computer-use agents (CUAs) capable of automating diverse computer tasks. As their commercial potential grows, critical details of the most capable CUA systems remain closed. As these agents will increasingly mediate digital interactions and execute consequential decisions on our behalf, the research community needs access to open CUA frameworks to study their capabilities, limitations, and risks. To bridge this gap, we propose OpenCUA, a comprehensive open-source framework for scaling CUA data and foundation models. Our framework consists of: (1) an annotation infrastructure that seamlessly captures human computer-use demonstrations; (2) AgentNet, the first large-scale computer-use task dataset spanning 3 operating systems and 200+ applications and websites; (3) a scalable pipeline that transforms demonstrations into state-action pairs with reflective long Chain-of-Thought reasoning that sustain robust performance gains as data scales. Our end-to-end agent models demonstrate strong performance across CUA benchmarks. In particular, OpenCUA-32B achieves an average success rate of 34.8% on OSWorld-Verified, establishing a new state-of-the-art (SOTA) among open-source models and surpassing OpenAI CUA (GPT-4o). Further analysis confirms that our approach generalizes well across domains and benefits significantly from increased test-time computation. We release our annotation tool, datasets, code, and models to build open foundations for further CUA research.
@article{wang2025opencua, title = {OpenCUA: Open Foundations for Computer-Use Agents}, author = {Wang, Xinyuan and Wang, Bowen and Lu, Dunjie and Yang, Junlin and Xie, Tianbao and Wang, Junli and Deng, Jiaqi and Guo, Xiaole and Xu, Yiheng and Wu, Chen Henry and Shen, Zhennan and Li, Zhuokai and Li, Ryan and Li, Xiaochuan and Chen, Junda and Zheng, Boyuan and Li, Peihang and Lei, Fangyu and Cao, Ruisheng and Fu, Yeqiao and Shin, Dongchan and Shin, Martin and Hu, Jiarui and Wang, Yuyan and Chen, Jixuan and Ye, Yuxiao and Zhang, Danyang and Du, Dikang and Hu, Hao and Chen, Huarong and Zhou, Zaida and Yao, Haotian and Chen, Ziwei and Gu, Qizheng and Wang, Yipu and Wang, Heng and Yang, Diyi and Zhong, Victor and Sung, Flood and Charles, Y. and Yang, Zhilin and Yu, Tao}, journal = {Advances in Neural Information Processing Systems}, year = {2025}, eprint = {2508.09123}, url = {https://arxiv.org/abs/2508.09123}, note = {NeurIPS 2025 Spotlight; Best Paper Award, COLM 2025 AIA Workshop} }

arXiv 2026

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

Zhiqin Yang, Chenxin Li, Xiaomeng Hu, Yibin Liu, Weidong Huang, … Xinyuan Wang

arXiv PDF

@misc{yang2026robotworld,
  title = {RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments},
  author = {Yang, Zhiqin and Li, Chenxin and Hu, Xiaomeng and Liu, Yibin and Huang, Weidong and Sun, Jiankai and Li, Haitao and Wu, Zijian and Huang, Yuzhi and Huang, Fanding and Sun, Hanwen and Liu, Jiashun and Tong, Jingqi and Huang, Mingxin and Hu, Shaoli and Huang, Shijue and Bai, Tianyi and Wang, Xinyuan and Lin, Yunlong and Tang, Zhengyang and Zhang, Zhexin and Chen, Zhuo and Song, Xierui and Dai, Juntao and Chen, Boyuan and Ji, Jiaming and Zhan, Fangneng and Hu, Mengkang and Xue, Wei and Zhang, Yonggang and Hu, Han and Ho, Tsung-Yi and Guo, Yike},
  year = {2026},
  eprint = {2610.10409},
  archiveprefix = {arXiv},
  url = {https://arxiv.org/abs/2610.10409}
}

Model release 2026

Hy4 preview

Tencent Hy Team

Code Website

@misc{tencent2026hy4,
  title = {Hy4 preview},
  author = {{Tencent Hy Team}},
  year = {2026},
  url = {https://huggingface.co/tencent/Hy4-preview}
}

Technical report 2026

Kimi K3: Open Frontier Intelligence

Kimi Team

arXiv PDF

@misc{kimi2026k3,
  title = {Kimi K3: Open Frontier Intelligence},
  author = {{Kimi Team}},
  year = {2026},
  eprint = {2607.24653},
  archiveprefix = {arXiv},
  url = {https://arxiv.org/abs/2607.24653}
}

arXiv 2026

OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks

Mengqi Yuan, Zilong Zhou, Xinzhuang Xiong, Weiming Wu, Jiayang Sun, … Xinyuan Wang

arXiv PDF

@misc{yuan2026osworld2,
  title = {OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks},
  author = {Yuan, Mengqi and Zhou, Zilong and Xiong, Xinzhuang and Wu, Weiming and Sun, Jiayang and Song, Jiamin and Cui, Kaiqian and Wang, Bowen and Wu, Haoyuan and Li, Yitong and Lu, Dunjie and Lu, Haikong and Zhen, Qi and Wang, Xinyuan and Deng, Jiaqi and Yang, Yuhao and Chen, Cheng and Zheng, Boyuan and Su, Alex and Yu, Xiao and Zou, Hao and Agashe, Saaket and Lu, Xing Han and Kaur, Manpreet and Qi, Zhengyang and Chen, Vincent Sunn and Sala, Frederic and Liu, Dayiheng and Lin, Junyang and Yu, Zhou and Su, Yu and Reddy, Siva and Wang, Xin Eric and Qi, Peng and Xie, Tianbao and Yu, Tao},
  year = {2026},
  eprint = {2606.29537},
  archiveprefix = {arXiv},
  url = {https://arxiv.org/abs/2606.29537}
}

arXiv 2026

PhoneBuddy: Training Open Models for Agentic Phone Use

Zhengyang Tang, Xin Lai, Pengyuan Lyu, Xinyuan Wang, Tianyi Bai

arXiv PDF Website

Phones are becoming an important execution surface for general-purpose agents, but training open models for reliable phone use remains difficult because the deployment environment is slow, stateful, side-effectful, and hard to reset or verify, while scalable mock environments only approximate real behavior. We present PhoneBuddy, a training recipe and open-model line that combines a real-app environment with a mock-app environment, PhoneWorld, which reconstructs runnable mock apps from real GUI usage structure. PhoneBuddy first builds a shared supervised fine-tuning stage, then compares real-app RL against mixed RL across both environments. On a 150-task human evaluation, success improves from 36.67% (SFT) to 40.67% (real-app RL) to 45.33% (mixed RL); on AndroidWorld, from 60.3% to 77.2% to 83.2%. Mock-app training is not a replacement for real-app RL, but a complementary source of scalable, resettable, and automatically checked interaction.
@misc{tang2026phonebuddy, title = {PhoneBuddy: Training Open Models for Agentic Phone Use}, author = {Tang, Zhengyang and Lai, Xin and Lyu, Pengyuan and Wang, Xinyuan and Bai, Tianyi and Li, Chenxin and Guo, Yiduo and Shen, Huawen and Liu, Yuxuan and Li, Junyi and Fang, Zhengyao and Ding, Yang and Zhang, Yi and Wang, Weinong and Zhou, Xingran and Wu, Liang and Tang, Fei and Fan, Sunqi and Peng, Shangpin and Ruan, Zheng and Zhang, Anran and Wang, Benyou and Wen, Ji-Rong and Yan, Rui and Zhang, Chengquan and Hu, Han}, year = {2026}, eprint = {2606.23049}, url = {https://arxiv.org/abs/2606.23049}, archiveprefix = {arXiv}, primaryclass = {cs.CL} }
arXiv 2026

GameCraft-Bench: Can Agents Build Playable Games End-to-End in a Real Game Engine?

Tongxu Luo, Rongsheng Wang, Jiaxi Bi, Chenming Xu, Zhengyang Tang, … Xinyuan Wang

arXiv PDF Website

We formalize end-to-end game generation, where coding agents must turn natural-language specifications into playable games within a game engine. Unlike traditional coding, scripts, scenes, assets, rendering, and runtime interactions must jointly produce coherent gameplay. We propose three desiderata—Engine Grounding, Artifact Completeness, and Interactive Verification—and introduce an evaluation framework using replayed demonstrations and rubric-guided multimodal judging. This is instantiated as GameCraft-Bench, a benchmark comprising 140 Godot tasks across 15 game families. The task remains difficult: the top agent reaches only 41.46%, with most below 40%. Agents implement recognizable mechanics but struggle with complete content, visual feedback, and presentation.
@misc{luo2026gamecraft, title = {GameCraft-Bench: Can Agents Build Playable Games End-to-End in a Real Game Engine?}, author = {Luo, Tongxu and Wang, Rongsheng and Bi, Jiaxi and Xu, Chenming and Tang, Zhengyang and Chen, Jianlong and Liang, Juhao and Ji, Ke and Guo, Shuqi and Du, Yuhao and Bu, Fan and Du, Wenyu and Zhang, Xiaotong and Li, Kyle and Wang, Shaobo and Zhang, Linfeng and Liu, Yuxuan and Lai, Xin and Li, Chenxin and Guo, Yiduo and Zhang, Zhexin and Wang, Xinyuan and Bai, Tianyi and Li, Ziniu and Wang, Benyou}, year = {2026}, eprint = {2606.17861}, url = {https://arxiv.org/abs/2606.17861}, archiveprefix = {arXiv}, primaryclass = {cs.CL} }
arXiv 2026

Do Phone-Use Agents Respect Your Privacy?

Zhengyang Tang, Ke Ji, Xidong Wang, Zihan Ye, Xinyuan Wang

arXiv PDF Code

We investigate whether phone-use agents protect user privacy when carrying out routine mobile tasks. This has been hard to assess because privacy-compliant behavior was not well-defined for such agents, and standard apps do not expose what data agents enter into forms. We introduce MyPhoneBench, a verifiable evaluation framework for privacy behavior in mobile agents. It defines privacy-respecting phone use through three principles—permissioned access, minimal disclosure, and user-controlled memory—via a privacy contract called iMy, combined with instrumented mock apps and rule-based auditing. Evaluating five frontier models across 10 mobile apps and 300 tasks, we find task success, privacy-compliant completion, and reuse of saved preferences are separate capabilities, with no single model excelling at all three. The most common failure is poor data minimization. Privacy issues stem from over-helpful execution of benign tasks, and evaluating success alone overstates how deployment-ready these agents are.
@misc{tang2026privacy, title = {Do Phone-Use Agents Respect Your Privacy?}, author = {Tang, Zhengyang and Ji, Ke and Wang, Xidong and Ye, Zihan and Wang, Xinyuan and Guo, Yiduo and Li, Ziniu and Li, Chenxin and Hu, Jingyuan and Chen, Shunian and Luo, Tongxu and Bi, Jiaxi and Qin, Zeyu and Wang, Shaobo and Lai, Xin and Lyu, Pengyuan and Li, Junyi and Xu, Can and Zhang, Chengquan and Hu, Han and Yan, Ming and Wang, Benyou}, year = {2026}, eprint = {2604.00986}, url = {https://arxiv.org/abs/2604.00986}, archiveprefix = {arXiv}, primaryclass = {cs.CR} }

Technical report 2026

Kimi K2.5: Visual Agentic Intelligence

Kimi Team

arXiv PDF

@misc{kimi2026k25,
  title = {Kimi K2.5: Visual Agentic Intelligence},
  author = {{Kimi Team}},
  year = {2026},
  eprint = {2602.02276},
  archiveprefix = {arXiv},
  url = {https://arxiv.org/abs/2602.02276}
}

ICLR 2026

VideoAgentTrek: Computer-Use Pretraining from Unlabeled Videos

Dunjie Lu, Yiheng Xu, Junli Wang, Haoyuan Wu, Xinyuan Wang

arXiv PDF OpenReview Code Website

@inproceedings{lu2025videoagenttrekcomputerusepretraining,
  title = {VideoAgentTrek: Computer-Use Pretraining from Unlabeled Videos},
  author = {Lu, Dunjie and Xu, Yiheng and Wang, Junli and Wu, Haoyuan and Wang, Xinyuan and Wang, Zekun and Yang, Junlin and SU, Hongjin and Chen, Jixuan and Chen, Junda and Mao, Yuchen and Lin, Junyang and Hui, Binyuan and Yu, Tao},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year = {2026},
  eprint = {2510.19488},
  url = {https://openreview.net/forum?id=xxYPqm1qWz},
  archiveprefix = {arXiv},
  primaryclass = {cs.CL}
}

ICLR 2026

Computer Agent Arena: Toward Human-Centric Evaluation and Analysis of Computer-Use Agents

Bowen Wang, Xinyuan Wang, Jiaqi Deng, Tianbao Xie, Ryan Li

PDF OpenReview Website

@inproceedings{wang2025computer,
  title = {Computer Agent Arena: Toward Human-Centric Evaluation and Analysis of Computer-Use Agents},
  author = {Wang, Bowen and Wang, Xinyuan and Deng, Jiaqi and Xie, Tianbao and Li, Ryan and Zhang, Yanzhe and Wang, Junli and Lu, Dunjie and Gong, Zicheng and Li, Gavin and Hua, Toh Jing and Chiang, Wei-Lin and Stoica, Ion and Yang, Diyi and Su, Yu and Zhang, Yi and Wang, Zhiguo and Zhong, Victor and Yu, Tao},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=3x4SDbXbgl}
}