Monday, 26 October 2026 · 6 min read
HiSTEVE integrates hierarchical planning and native control for Minecraft agents
HiSTEVE combines a capability-aware harness with a vision-language-action policy to achieve 85.0% atomic task success on MineStudio Bench and 37.4% average success on OpenHA, enabling long-horizon tasks like crafting an iron pickaxe from scratch.
Based on the paper HiSTEVE 1.0: Harnessing Human-Interface VLA Agents for Open-World Games by Zecong Tang, Zian Guan, Yuncheng Guo, Zhanqiu Zhang and 2 others
Challenges in building long-horizon native-interface agents for open-world games
Creating virtual embodied agents that operate in open-ended games like Minecraft requires two key capabilities: long-horizon autonomy to maintain and adapt plans over extended tasks, and native-interface control to translate first-person visual observations into keyboard-and-mouse actions reliably.
Existing agents typically excel at either native visuomotor control without long-horizon autonomy or high-level planning executed through environment APIs or predefined skills, but not both simultaneously.
Integrating these capabilities is challenging because long-horizon planning demands persistent task-state tracking and adaptive decision-making, while native-interface control must handle heterogeneous interactions such as GUI and non-GUI tasks, which differ in visual input and mouse-action semantics.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short