Researchers have developed ProCUA-SFT, a new dataset designed to improve the training of computer-use agents (CUAs) that interact with graphical desktop environments. Existing datasets like AgentNet have shown negative transfer effects, hindering performance. ProCUA-SFT, comprising 3.1 million step-level samples from synthetic trajectories, addresses this by using an automated pipeline for task generation and verification. Fine-tuning the UI-TARS 7B model on ProCUA-SFT resulted in a significant performance increase on the OSWorld benchmark, outperforming models trained on AgentNet. A portion of ProCUA-SFT was also integrated into the Nemotron 3 Nano Omni model to enhance its computer-use capabilities. AI
IMPACT This new dataset significantly improves AI agents' ability to interact with desktop environments, potentially accelerating the development of more capable and autonomous software agents.
RANK_REASON The cluster describes a new dataset and technical report detailing its creation and performance improvements for AI agents, which falls under research.
Read on Hugging Face Daily Papers →
- AgentNet
- Kimi K2.5
- Nemotron 3 Nano Omni
- OSWorld
- ProCUA-SFT
- SpreadsheetBench
- UI-TARS 7B
- Zenodo10K
- arXiv
- Hugging Face
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
