Reproduce “ISE: An Execution-Grounded Recipe for Multi-Turn OS-Agent Trajectories” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Siyuan Luo, Nairong Zheng, Lin Zhou, Tiankuo Yao, Shengyou Yuan, Haojia Yu, Cong Pang, Jiapeng Luo and 1 more
ISE creates multi-turn OS-agent training data by sampling structured user intents, keeping a simulated user role-locked, and executing tool calls in live isolated workspaces. Follow-up requests respond to actual outcomes. Its corpus contains 43,956 intents and 23,132 trajectories. After fine-tuning, Qwen3-8B reaches 37.7% ClawEval pass@1 versus 19.3% zero-shot; this result covers 114 real-shell tasks, not GUI or browser agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.