Junli Wang Roughly: “wahng jwin-lee”
Ph.D. Student · UC San Diego
I am a Ph.D. student at UC San Diego, advised by Prof. Prithviraj Ammanabrolu and Prof. Hao Zhang. My research focuses on building multimodal digital agents that can perceive, reason, and execute long-horizon tasks across diverse interfaces. I work on:
- Foundation models for digital agents, e.g. [Aguvis], [OpenCUA] and [Mocha-Coder].
- Scaling digital agent data through large-scale synthetic data mined from the internet, e.g. [AgentTrek], [VideoAgentTrek], and [Jedi].
- Evaluating digital agents in the wild, e.g. [OSWorld-Verified], [Computer Agent Arena], and [CocoaBench].
- Open-source infrastructure that makes agent rollout and learning practical at scale, such as [NanoRollout].
Beyond my own projects, I have contributed to post-training open-source foundation models for agentic capabilities, including [Qwen3.5] and [Qwen3-Coder]. I mostly work on agentic RL infrastructure that enables agent rollouts in large-scale digital environments and SFT data curation for digital agents.
Experience
-
UC San Diego Jan 2026 – PresentPh.D., Computer Science and EngineeringAdvised by Prof. Prithviraj Ammanabrolu and Prof. Hao Zhang
-
Tsinghua University Sep 2021 – Jun 2025B.Eng., Computer Science and TechnologyAffiliated with Xinya CollegeXinya College is the first residential undergraduate liberal arts and sciences college established by Tsinghua University on September 27, 2014. Admitting a select cohort of top students across the arts and sciences each year, it offers a year of broad liberal education before students choose any major at Tsinghua.
Models
Publications
-
TL;DR: VideoAgentTrek mines screen-recorded internet videos into labeled computer-use actions through inverse dynamics, creating web-scale supervision that improves agent pretraining without manual trajectory annotation.International Conference on Learning Representations (ICLR), 2026.
-
TL;DR: CocoaBench evaluates unified digital agents on long-horizon tasks that require composing vision, search, and coding, showing that current systems remain far from reliable in the wild.
-
TL;DR: Computer Agent Arena is an open-source platform for head-to-head CUA evaluation and a dynamic methodology that converts human preferences into structured feedback in realistic environments.International Conference on Learning Representations (ICLR), 2026.
-
TL;DR: OSWorld-Verified upgrades OSWorld with repaired tasks, more robust infrastructure, and public verified evaluation to provide more reliable signals for computer-use agent benchmarking.
-
TL;DR: OpenCUA releases an open framework for computer-use agents, including demonstration-capture infrastructure, AgentNet data, reasoning-augmented training pipelines, and competitive open-source CUA models.Neural Information Processing Systems (NeurIPS), 2025.
-
TL;DR: Aguvis builds a unified vision-based GUI agent that operates directly on screenshots with a standardized action space and structured reasoning, achieving strong cross-platform autonomous computer-use performance.International Conference on Machine Learning (ICML), 2025.
-
TL;DR: AgentTrek turns web tutorials into verified multimodal GUI agent trajectories, providing a scalable and low-cost alternative to human annotation for training stronger web agents.International Conference on Learning Representations (ICLR), 2025.
Powered by Jekyll and Minimal Light theme.
