Reproduce “Terminal-World: Scaling Terminal-Agent Environments via Agent Skills” on OpenResearch

About this paper

Zihao Cheng, Hongru Wang, Zeming Liu, Xinyi Wang, Xiangrong Zhu, Yuhang Guo, Wei Lin, Jeff Z. Pan and 1 more

Terminal-World introduces an automated pipeline that synthesizes high-quality training data for terminal-based AI agents by leveraging human-authored 'agent skills' from open-source ecosystems. Models trained on this skill-grounded data consistently outperform existing baselines across various benchmarks, achieving higher sample efficiency while requiring significantly less training data.

Key insights

• Skill-grounded trajectories, even when explicit guidelines are removed during Supervised Fine-Tuning (SFT), enable agents to learn efficient, reliable autonomous planning, and even outperform their teacher models in terms of steps and command failure rates.
• Including 'failed' trajectories in the training data is crucial for performance, as these trajectories often contain valuable intermediate correct steps and expose error-correction mechanisms, contributing more to learning than just successful paths.
• User personas serve as an effective context multiplier, significantly enhancing task diversity and scenario coverage by instantiating the same underlying skills across varied user backgrounds and goals.

Summary by alphaXiv.