Overview

ASIWorld is a persistent survival sandbox for studying how autonomous agents carry out goals over extended interactions. Agents work on interdependent production and exploration tasks using character-scoped observations, direct actions, and programs that run asynchronously between model calls. Evaluation combines task completion, development achievements, and survival costs with analysis of execution behavior, revealing differences in preparation, combat, and recovery even among agents with similar task outcomes.

BibTeX

@article{lin2026asiworld,
  title={ASIWorld: Benchmarking Autonomous Agents in a Persistent Survival Sandbox},
  author={Lin, Jianfeng and Wang, Runze and Hong, Haorong and Xie, Yuxuan and Zeng, Zeno and Liu, Zhongyuan},
  journal={arXiv preprint arXiv:2610.14807},
  year={2026}
}