Bowen Zhou

Bowen Zhou

M.S. Student in Artificial Intelligence · Tsinghua University · Research Intern @ Qwen Team, Alibaba

Building agents that operate GUIs

Bowen Zhou

I am an M.S. student in Artificial Intelligence at Tsinghua University. Previously, I received my B.Eng. in Artificial Intelligence from Dalian University of Technology.

My research focuses on building Universal Digital Agents (UDAs), with an emphasis on Computer-Use Agents. I study how agents can understand and operate across digital environments, including graphical interfaces, terminals, and code. I am particularly interested in long-horizon reasoning, post-training, and Agentic RL for UDAs.

I aim to make computer use a core capability of frontier models and to bring UDAs into everyday life. Computers serve as the core medium for these agents. By operating digital systems and connecting them with real-world tasks, they can improve accessibility, extend human capabilities, and create meaningful impact in the physical world. I believe UDAs represent an important path toward AGI.

I am currently a research intern with the Qwen Team at Alibaba, where I am advised by Que Shen. Previously, I interned at Huawei Multimodal Lab and collaborated remotely with Microsoft Research Asia.

If you would like to talk about research or life, feel free to reach out. I am open to research collaborations and opportunities.

news

  1. Released RecreationWorld, scalable and verifiable environments for hybrid Computer-Use Agents. new
  2. WeaveBench and ST-Lite accepted by EMNLP 2026.
  3. LoopX reached #1 on GitHub Trending.
  4. Started a research internship with the Qwen Team at Alibaba.
  5. Released WeaveBench, a benchmark for hybrid-interface Computer-Use Agents.
  6. Started a research internship at Huawei Multimodal Lab.
  7. Released ST-Lite for efficient long-horizon GUI agents.
  8. Began a remote research collaboration with Microsoft Research Asia.

selected publications

Toward universal agents that can perceive, reason, and act across diverse digital interfaces.

  1. Original application
    Recreated application
    Original app Recreation
    Food Truck · macOS · recreated by GPT-6 Astra
    Tech Report
    RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents
    Core Contributor
    Technical Report, 2026

    A framework that turns software recreation into scalable training environments with verifiable rewards, alongside a 250-task benchmark across Ubuntu, macOS, Windows, Android, and Web.

    PDF arXiv Website Code Dataset Figure
  2. SPA · Audit a 3D mesh for printability in MeshLab
    EMNLP 2026
    WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces
    First Author
    Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026

    A benchmark of 114 real-world tasks across eight domains that require agents to interleave GUI observation with CLI and code execution; the best evaluated system reaches only 41.2% success.

    arXiv Website Code Dataset
    1. Method. ST-Lite keeps the current step in full and only the salient, still-attended tokens of earlier screenshots, holding the KV cache at a fixed 20% budget. Schematic.
      EMNLP 2026
      ST-Lite: Training-Free KV Cache Compression with Spatio-Trajectory Guidance for Long-Horizon GUI Agents
      First Author
      Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026

      A training-free KV cache compression method tailored to GUI interaction traces, matching or exceeding Full Cache accuracy at a 20% budget while delivering up to 2.35× decoding speedup.

      arXiv Code Figure
    2. Method. Older frames are encoded at lower resolution, and flat background patches of the current screenshot are pruned by UI structure. Schematic; real screenshot from RecreationBench.
      arXiv 2026
      GUIPruner: Spatio-Temporal Token Pruning for Efficient High-Resolution GUI Agents
      First Author
      arXiv preprint, 2026

      A training-free framework that combines temporal-adaptive resolution with structure-aware visual-token pruning, achieving a 3.4× FLOPs reduction and a 3.3× vision-encoding speedup while retaining over 94% of the original performance.

      arXiv Figure

open source

LoopX ★ 5,923 · 563 forks
Membership · Long-Horizon Agent Control Plane · #1 on GitHub Trending, Aug 5, 2026

An open-source framework for long-horizon agents that preserves task state across sessions and coordinates work across Codex, Claude Code, and other agent tools. It supports recovery, handoffs, and human oversight so agents can continue complex work reliably.

Website Code Docs
Demo. The LoopX workspace turns scattered agent sessions into a goal board; when a new constraint arrives, the goal replans and evidence closes the loop. Clips from the official LoopX demos, sped up.

experience

education

skills

ResearchUniversal Digital Agents · Agentic RL · Multimodal Foundation Models · Inference Optimization
LanguagesPython · C/C++ · LaTeX
FrameworksPyTorch · veRL · slime · vLLM · SGLang · Transformers
SystemsDocker · Kubernetes · VM-based agent environments · large-scale evaluation