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

arXiv cs.CL · 18 hours ago

Researchers introduced ISE, a three-stage method for generating training data for OS agents that creates structured user intents, simulates multi-turn interactions, and executes tool calls in live environments. The dataset contains 43,956 unique intents and 23,132 trajectories averaging 8.12 user turns each, with generated data improving a Qwen3-8B model's performance on ClawEval from 19.3 to 37.7 pass@1. This approach allows smaller models to outperform zero-shot GPT-4o on agent tool-use tasks through better training data grounded in actual execution outcomes.