KPHarness: A Game AI Execution Framework with Co-evolution of Knowledge and Probing
What happened
Zhikang Chen proposes KPHarness, a method that adaptively adjusts the external execution harness of a frozen vision-language-action (VLA) model through the co-evolution of knowledge and probing, for long-horizon game tasks. The work appears in an arXiv paper. The paper says that on Minecraft mining, combat, and crafting tasks, KPHarness achieves higher task coverage than pure policy execution, a manually fixed harness, and an approach that evolves only knowledge without adaptive probing. On unseen tasks, keeping the evolved harness, knowledge base, and probing strategy fixed also improves task completion rate. These results are self-reported and have not been independently verified.
AI-written from the coverage · updated 2026-10-21
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Game AIKPHarness: Co-evolving knowledge and probing for long-horizon gaming harnesses
KPHarness adjusts the external execution harness of a frozen vision-language-action (VLA) model through co-evolution of knowledge and probing. On Minecraft mining, combat, and crafting tasks, its task coverage beats pure policy execution, manually fixed harnesses, and a scheme that evolves only knowledge with non-adaptive probing. On unseen tasks, task completion also improves when the evolved harness, knowledge base, and probing strategy are kept fixed.