| CARVIEW |
Select Language
HTTP/2 200
server: GitHub.com
content-type: text/html; charset=utf-8
last-modified: Wed, 24 Dec 2025 07:55:21 GMT
access-control-allow-origin: *
strict-transport-security: max-age=31556952
etag: W/"694b9c69-222a5"
expires: Sun, 28 Dec 2025 13:20:58 GMT
cache-control: max-age=600
content-encoding: gzip
x-proxy-cache: MISS
x-github-request-id: 4FD1:2C10E1:7A7C61:894667:69512C61
accept-ranges: bytes
date: Sun, 28 Dec 2025 13:54:53 GMT
via: 1.1 varnish
age: 0
x-served-by: cache-bom-vanm7210090-BOM
x-cache: HIT
x-cache-hits: 0
x-timer: S1766930093.854792,VS0,VE201
vary: Accept-Encoding
x-fastly-request-id: a8e9bb45636a9dd5d36b01758267513aa73236db
content-length: 29182
OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
**OSWorld** is a first-of-its-kind scalable, real computer environment for multimodal agents,
supporting task setup, execution-based evaluation, and interactive learning across operating systems.
It can serve as a unified environment for evaluating open-ended computer tasks that involve arbitrary
apps (e.g., task examples in the above Fig). We also create a benchmark of 369 real-world computer
tasks in **OSWorld** with reliable, reproducible setup and evaluation scripts. *Note: 8 Google Drive tasks may require manual configuration or can be excluded (361 tasks) due to network dependencies.*
Autonomous agents that accomplish complex computer tasks with minimal human
interventions have the potential to transform human-computer interaction, significantly enhancing
accessibility and productivity. However, existing benchmarks
either lack an interactive environment or are limited to environments specific to
certain applications or domains, failing to reflect the diverse and complex nature of real-world computer
use, thereby limiting the scope of tasks and agent
scalability. To address this issue, we introduce **OSWorld**, the first-of-its-kind
scalable, real computer environment for multimodal agents, supporting task setup,
execution-based evaluation, and interactive learning across various operating systems such as Ubuntu,
Windows, and macOS. **OSWorld** can serve as a unified,
integrated computer environment for assessing open-ended computer tasks that
involve arbitrary applications. Building upon **OSWorld**, we create a benchmark
of 369 computer tasks involving real web and desktop apps in open domains, OS
file I/O, and workflows spanning multiple applications (note: 8 Google Drive tasks may require manual setup or can be excluded for a 361-task evaluation). Each task example is
derived from real-world computer use cases and includes a detailed initial state
setup configuration and a custom execution-based evaluation script for reliable,
reproducible evaluation. Extensive evaluation of state-of-the-art LLM/VLM-based
agents on **OSWorld** reveals significant deficiencies in their ability to serve as
computer assistants. While humans can accomplish over 72.36% of the tasks, the
best model achieves only 12.24% success, primarily struggling with GUI grounding
and operational knowledge. Comprehensive analysis using **OSWorld** provides
valuable insights for developing multimodal generalist agents that were not possible with previous
benchmarks.
The **OSWorld** environment uses a configuration file for
initializing tasks *(highlighted in red)*, agent interaction, post-processing upon agent
completion *(highlighted in orange)*, retrieving files and information *(highlighted in yellow)*, and
executing the evaluation function *(highlighted in green)*. The corresponding configuration items are
highlighted in colors that match their respective components within the environment. Environments
can run in parallel on a single host machine for learning or evaluation purposes. Headless operation
is supported.
Below we present an overview of the main statistics of **OSWorld**, showcasing the outline and a broad spectrum
of tasks. **OSWorld** contains a total of 369 tasks (and an additional 43 tasks on Windows for
analysis).
Key statistics of
OSWorld.
The "Supp. tasks" refers to the Windows-based tasks, that could
only be used after
activation due to copyright restrictions.
We make a comparison of **OSWorld** against some other different benchmarks for digital agents as
presented below.
**The columns indicate:** whether they provide a controllable executable environment *(Control. Exec. Env.)*, the ease of adding new tasks involving arbitrary applications in open domains *(Environment Scalability)*, support for multimodal agent evaluation *(Multimodal Support)*, support for and inclusion of cross-app tasks *(Cross-App)*, capability to start tasks from an intermediate initial state *(Intermediate Init. State)*, and the number of execution-based evaluation functions *(# Exec.-based Eval. Func.)*.
We adopt state-of-the-art LLM and VLM from open-source representatives such as UI-TARS, Agent-S, Qwen, Mixtral and
CogAgent, and closed-source ones from Operator, GPT, Gemini, and Claude families on **OSWorld**, as LLM and VLM agent
baselines. We also explore methods such as the Set-of-Marks aided
approach, which has been demonstrated to improve spatial capabilities for visual
reasoning.
**We are actively updating the benchmark with new LLMs, VLMs and methods. Pull requests welcomed!**
What are the differences among General model, Specialized model, and Agentic framework?
A General model is a model with broad, general-purpose capabilities. "Computer use" is one capability that can be elicited via prompting; the model itself can still perform other tasks such as dialogue and code generation. A Specialized model is trained specifically to serve as a computer-use agent; other capabilities are out of scope and are not emphasized in the corresponding reports. An Agentic framework organizes one or more General and Specialized models into a structured workflow—commonly, a GPT-family model acts as the planner while a proprietary or task-specific model serves as the grounder.
We will add new paradigms as they emerge.
We conduct a qualitative analysis in the aspect of models, methods, and human to find out factors influencing
the
performance of VLMs in digital agent tasks and their underlying behavioral logic. We investigate the impact of
task attributes
(such as difficulty, feasibility, visual requirement, and GUI complexity), input measurements
(such as screenshot resolution, the influence of trajectory history, and the effect of UI layout), and
explore
whether there are patterns in the agent's performance across different operating systems. Here is an overview of
our analysis outcome.
Higher screenshot resolution typically leads to improved performance.
Longer text-based trajectory history context improves performance, unlike screenshot-only history,
but poses efficiency challenges.
Current VLM agents are not robust to UI layout and noise.
The performance of VLM agents across different OS is in strong correlation. This implies that
insights and
methodologies developed within the OSWorld framework can be effectively transferred to
Windows environments
with a high degree of reliability.
A success case of LLM/VLM agent baselines.
Special thanks to the following YouTubers and enthusiasts for their reports. We are delighted to see the
community's interest. If you would like a brief video introduction and their thoughts, feel free to check
them out!
Special thanks to the following institutions that provided feedback and participated in the fixes (as well as institutions that provided feedback during the process): [MoonShot AI, a.k.a. Kimi](https://www.moonshot.ai/)，[Human Data](https://www.hud.so/), [OpenAI](https://openai.com/), [ByteDance Seed TARS](https://seed-tars.com/), [Anthropic](https://www.anthropic.com/), [Simular](https://www.simular.ai/), [HKU Data Intelligence Lab](https://sites.google.com/view/chaoh)
Special thanks to the following students who participated in the specific fixes: [Mengqi Yuan](https://yuanmengqi.github.io/), [Danyang Zhang](https://zdy023.github.io/), [Xinzhuang Xiong](https://thisisxxz.com/), [Zhennan Shen](https://scholar.google.com/citations?user=JPwg5MwAAAAJ&hl=en), [Zilong Zhou](https://github.com/adlsdztony), Yanxu Chen, [Jiaqi Deng](https://millank0817.github.io/), [Tianbao Xie](https://tianbaoxie.com/), Junda Chen, [Jixuan Chen](https://chenjix.github.io/), [Haoyuan Wu](https://www.linkedin.com/in/haoyuan-wu-240878291/).
Special thanks to the following students who participated in running the re-evaluation: [Mengqi Yuan](https://yuanmengqi.github.io/), [Zilong Zhou](https://github.com/adlsdztony), [Xinyuan Wang](https://xinyuanwangcs.github.io/), [Bowen Wang](https://bowenbryanwang.github.io/).
Please start by reading through the [agent interface](https://github.com/xlang-ai/OSWorld/blob/main/mm_agents/README.md) and the [environment interface](https://github.com/xlang-ai/OSWorld/blob/main/desktop_env/README.md).
Correctly implement the agent interface and import your customized version in the `run.py` or `run_multienv.py` file.
Afterward, you can execute a command similar to the one in the previous section to run the benchmark on your agent.
If you want your results to be verified and displayed on the verified leaderboard, you need to schedule a meeting with us (current maintainer: tianbaoxiexxx@gmail.com, yuanmengqi732@gmail.com) to run your agent code on our side and have us report the results.
You need to upload and allow us to disclose your agent implementation under the OSWorld framework (you may choose not to expose your model API to the public), along with a report that allows the public to understand what's happening behind the scenes.
Alternatively, if you are from a trusted institution, you can share your monitoring data and trajectories with us.
Please carefully follow the [Public Evaluation Guideline](https://github.com/xlang-ai/OSWorld/blob/main/PUBLIC_EVALUATION_GUIDELINE.md) to get results.
Option 1 - Manual Configuration: Manually troubleshoot and configure these 8 tasks to complete the full 369-task evaluation.
Option 2 - Task Exclusion: Exclude these 8 tasks and run the remaining 361 tasks - this is officially permitted and acceptable for benchmark evaluation.
Both approaches are valid for research comparison and leaderboard submission. Please specify which approach you used when reporting your results.
OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
Tianbao Xie1,
Danyang Zhang1,
Jixuan Chen1,
Xiaochuan Li1,
Siheng Zhao1, Ruisheng Cao1, Toh Jing Hua1, Zhoujun Cheng1, Dongchan Shin1, Fangyu Lei1, Yitao Liu1, Yiheng Xu1, Shuyan Zhou3, Silvio Savarese2, Caiming Xiong2, Victor Zhong4, Tao Yu1
Siheng Zhao1, Ruisheng Cao1, Toh Jing Hua1, Zhoujun Cheng1, Dongchan Shin1, Fangyu Lei1, Yitao Liu1, Yiheng Xu1, Shuyan Zhou3, Silvio Savarese2, Caiming Xiong2, Victor Zhong4, Tao Yu1
1The University of Hong Kong,
2Salesforce Research,
3Carnegie Mellon University,
4University of Waterloo
2025-07-28: Major Upgrade! OSWorld has been enhanced and is now OSWorld-Verified with comprehensive improvements: fixed community-reported examples, AWS support reducing evaluation time to within 1 hour, and updated benchmark results. See the verified benchmark results in the Benchmark section below. Please compare your OSWorld results with the new benchmark results when running the latest version.
Abstract
OSWorld Environment Infrastructure
Data Statistics and Comparison
Key statistics of
OSWorld.
The "Supp. tasks" refers to the Windows-based tasks, that could
only be used after
activation due to copyright restrictions.
Distribution of task instructions in OSWorld
based on the app domains and operation
types to showcase the content intuitively.
**The columns indicate:** whether they provide a controllable executable environment *(Control. Exec. Env.)*, the ease of adding new tasks involving arbitrary applications in open domains *(Environment Scalability)*, support for multimodal agent evaluation *(Multimodal Support)*, support for and inclusion of cross-app tasks *(Cross-App)*, capability to start tasks from an intermediate initial state *(Intermediate Init. State)*, and the number of execution-based evaluation functions *(# Exec.-based Eval. Func.)*.
| OSWorld | |
|---|---|
| # Instances (# Templates) |
369 |
| Control. Exec. Env. |
Computer |
| Environment Scalability? | ✔️ |
| Multimodal Support? |
✔️ |
| Cross-App? | ✔️ |
| Intermediate Init. State? |
✔️ |
| # Exec.-based Eval. Func. |
134 |
| GAIA | Mind2Web | WebLINX | PixelHelp | MetaGUI | AitW | OmniAct | ScreenAgent | AgentBench | InterCode | MiniWoB++ | WebShop | WebArena | VisualWebArena | WorkArena | WikiHow | AssistGUI |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 466 |
2350 | 2337 | 187 | 1125 | 30k | 9802 | 70 | 1091 | 1350(3) | 125 | 12k(1) | 812(241) | 910(314) | 23k(29) | 150(16) | 100 |
| ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | Multi-isolated | Code | Web | Web | Web | Web | Web | Mobile | ❌ |
| - | - | - | - | - | - | - | - | ❌ |
❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| ❌ |
✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ❌ | ❌ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
| ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| ❌ | ✔️ | ✔️ | ❌ |
❌ | ✔️ | ✔️ | ✔️ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✔️ | ❌ | ✔️ |
| 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 7 | 3 | 125 | 1 | 5 | 6 | 7 | 16 | 2 |
Benchmark
Important Notice: Google Drive Tasks (2025-07-28)
OSWorld contains 8 Google Drive-related tasks that may encounter setup issues during task initialization due to IP changes or other network-related factors, even when following our configuration guidelines correctly.
Two acceptable approaches for evaluation:
- Manual Adjustment: You can manually configure these 8 tasks to complete the full 369 tasks evaluation
- Exclude Tasks: You can exclude these 8 tasks and run 361 tasks instead - this is officially permitted and acceptable
Both approaches are valid for benchmark comparison and leaderboard submission.
Results
These are official results evaluated by our team under unified settings and environment. All models are tested with consistent evaluation protocols to ensure fair comparison.
For self-reported results and progress trends across different modalities, click here.
All verified trajectories are hosted on Hugging Face for community analysis.
A General model is a model with broad, general-purpose capabilities. "Computer use" is one capability that can be elicited via prompting; the model itself can still perform other tasks such as dialogue and code generation. A Specialized model is trained specifically to serve as a computer-use agent; other capabilities are out of scope and are not emphasized in the corresponding reports. An Agentic framework organizes one or more General and Specialized models into a structured workflow—commonly, a GPT-family model acts as the planner while a proprietary or task-specific model serves as the grounder.
We will add new paradigms as they emerge.
Loading verified benchmark data...
Analysis
Videos
@Yannic Kilcher
@Wes Roth
@hu-po
@Dylan Curious
@WorldofAI
@Gourcer
@AI Explained
@Fireship
@1littlecoder
Acknowledgement
We thank Sida Wang, Peter Shaw, Alane Suhr, Luke Zettlemoyer, Haoyuan Wu, Junli Wang, Chengyou Jia, Junlin Yang, Junlei Zhang, Chen Henry Wu, Pengcheng Yin, Shunyu Yao, Xing Han Lu, Siva Reddy, Ruoxi Sun, Zhiyuan Zeng, and Lei Li for their helpful feedback on this work
Acknowledgement for OSWorld-Verified
Evaluation
Local Evaluation
Public Evaluation
FAQ
What is the username and password for the virtual machines?
The username and password for the virtual machines are as follows (for provider vmware, virtualbox and docker): we set the account credentials for Ubuntu as user / password.
For cloud service providers like aws, to prevent attacks due to weak passwords, we default to osworld-public-evaluation.
If you make further modifications, remember to set the client_password variable and pass it to DesktopEnv and Agent (if supported) when running experiments.
Some features like setting up proxy require the environment to have the client VM password to obtain sudo privileges, and for some OSWorld tasks, the agent needs the password to obtain sudo privileges to complete them.
How to setup the account and credentials for Google and Google Drive?
See Account Guideline.
What should I do if Google Drive tasks fail to initialize properly?
OSWorld contains 8 Google Drive-related tasks that may encounter setup issues during initialization due to various factors:
Common Issues:
- IP address changes causing authentication problems
- Network restrictions or firewalls
- Google API rate limiting or access restrictions
- Regional availability limitations
Option 1 - Manual Configuration: Manually troubleshoot and configure these 8 tasks to complete the full 369-task evaluation.
Option 2 - Task Exclusion: Exclude these 8 tasks and run the remaining 361 tasks - this is officially permitted and acceptable for benchmark evaluation.
Both approaches are valid for research comparison and leaderboard submission. Please specify which approach you used when reporting your results.
How can I configure a proxy for the VM (if I'm behind the GFW, or I don't want some of my tasks to be identified as bot and get lower scores)?
If you want to set it up yourself, please refer to Proxy Guideline.
We also provide a pre-configured solution based on dataimpulse, please refer to proxy-setup section in PUBLIC_EVALUATION_GUIDELINE.
BibTeX
@misc{OSWorld,
title={OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments},
author={Tianbao Xie and Danyang Zhang and Jixuan Chen and Xiaochuan Li and Siheng Zhao and Ruisheng Cao and Toh Jing Hua and Zhoujun Cheng and Dongchan Shin and Fangyu Lei and Yitao Liu and Yiheng Xu and Shuyan Zhou and Silvio Savarese and Caiming Xiong and Victor Zhong and Tao Yu},
year={2024},
eprint={2404.07972},
archivePrefix={arXiv},
primaryClass={cs.AI}
}