ISE: An Execution-Grounded Recipe for Multi-Turn OS-Agent Trajectories

The ISE (Intent -> Simulate -> Execute) framework has been introduced to enhance the training of operating system (OS) agents by generating a dataset that captures structured user intents and multi-turn task delegation. The framework synthesizes approximately 50,000 structured intents and produces 23,132 multi-turn interaction trajectories, achieving a notable improvement in performance metrics, such as a pass@1 score increase from 19.3 to 37.7 using Qwen3-8B on agent tool-use tasks. This release is significant for AI practitioners as it provides a comprehensive dataset and methodology to improve the grounded execution capabilities of OS agents, addressing a critical gap in existing resources.