Reproduce “MineEvolve: Self-Evolution with Accumulated Knowledge for Long-Horizon Embodied Minecraft Agents” 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
Zhengwei Xie, Zhisheng Chen, Ziyan Weng, Jinhan Li, Chenglong Li, Zikai Xiao, Jingwei Song, Jinhao Jing and 2 more
MineEvolve is a framework enabling embodied Minecraft agents to continuously improve on long-horizon tasks by converting fine-grained execution feedback into structured, actionable behavioral knowledge. It achieved the highest overall success rates across multiple large language model backbones, notably improving performance on complex tasks within a 70-task benchmark by systematically generating and applying skills and remedies.
Key insights
- Converting fine-grained, structured (typed) execution feedback into explicit behavioral knowledge (skills and remedies) is more effective for continuous self-improvement than free-form textual reflections or simple binary feedback.
- Targeted remedies for execution failures and stagnation are particularly crucial for enhancing agent robustness and enabling recovery in long-horizon, high-dependency tasks.
- A robust knowledge curation mechanism, including validation, merging, and conflict resolution, is essential for maintaining a high-quality and non-conflicting knowledge base to guide future actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.