OSCAR

Overview

OSCAR is an open-source agent system developed by Universite de Montreal & Mila that achieves strong performance using screenshot-based interaction.

Key Features

• Screenshot-based interaction
• GPT-4o integration
• Academic research focus
• Developed by Mila team

Performance

OSWorld Results

• OSCAR w/ GPT-4o: 24.5% (best screenshot-based model)

Technical Details

• Base Model: GPT-4o integration
• Input: Screenshot-based
• Focus: Visual understanding and interaction

References

• Paper: [Wang et al., '24]
• Citation: https://arxiv.org/pdf/2410.18963