OSWorld

January 23, 2025 ยท View on GitHub

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.

Key Features

  • Scalable computer environment supporting multiple operating systems (Ubuntu, Windows, macOS)
  • Unified environment for evaluating open-ended computer tasks
  • Supports arbitrary applications and cross-app workflows
  • Includes 369 real-world computer tasks with reliable setup and evaluation scripts
  • Enables interactive learning and evaluation

Performance Metrics

Latest benchmark results as of April 2024:

Overall Performance

  • Human Performance: >72.36% task success rate
  • Best Model Performance: 24.6% (UI-TARS-72B-DPO with 50 steps)
  • Models primarily struggle with GUI grounding and operational knowledge

Detailed Results by Model Type

A11y Tree Based Models

RankModelSuccess RateNotes
1UI-TARS-72B-DPO (50 steps)24.6%ByteDance & Tsinghua
2UI-TARS-72B-DPO (15 steps)22.7%ByteDance & Tsinghua
3Claude 3.5 Sonnet (50 steps)22.0%Anthropic
4UI-TARS-72B-SFT (15 steps)18.8%ByteDance & Tsinghua
5UI-TARS-7B-DPO (15 steps)18.7%ByteDance & Tsinghua

Screenshot Based Models

RankModelSuccess RateNotes
1OSCAR w/ GPT-4o24.5%Universite de Montreal
2GPT-4 Vision11.77%OpenAI
3GPT-4 Vision (0409)8.40%OpenAI
4Gemini-Pro-1.57.79%Google
5Claude-3-Opus6.72%Anthropic

Screenshot + A11y Tree Models

RankModelSuccess RateNotes
1Aguvis-72B w/ GPT-4o17.04%Salesforce & HKU
2Aria-UI w/ GPT-4o15.15%HKU & Rhymes AI
3OS-Atlas-Base-7B w/ GPT-4o14.63%Shanghai AI Lab
4Aguvis-72B10.26%Salesforce & HKU
5SeeClick w/ GPT-4o9.21%Nanjing University

Key Findings

  • Higher screenshot resolution improves performance
  • Longer text-based trajectory history helps but poses efficiency challenges
  • Current VLM agents are not robust to UI layout and noise
  • Performance across different OS shows strong correlation

Task Categories

Tasks span multiple domains including:

  • Web applications
  • Desktop applications
  • OS file I/O operations
  • Cross-application workflows
  • Real-world use cases

Technical Infrastructure

  • Uses configuration files for task initialization
  • Supports agent interaction and post-processing
  • Enables file and information retrieval
  • Includes execution-based evaluation functions
  • Supports parallel environment execution
  • Enables headless operation

Key Statistics

  • Total tasks: 369 (plus 43 additional Windows-based tasks)
  • Execution-based evaluation functions: 134
  • Supports intermediate initialization states
  • Enables multimodal interaction
  • Allows cross-app task execution

References