![[2505.24157] Experience-based Knowledge Correction for Robust Planning in Minecraft](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2505.24157] Experience-based Knowledge Correction for Robust Planning in Minecraft
Summary
The paper presents XENON, an advanced agent for robust planning in Minecraft that utilizes experience-based knowledge correction to improve long-horizon planning and knowledge acquisition.
Why It Matters
This research addresses the limitations of existing LLM-based agents in dynamic environments like Minecraft, where flawed initial knowledge can hinder performance. By introducing XENON, which learns from both successes and failures, the study enhances the potential for AI agents to adapt and improve over time, making it significant for advancements in AI and machine learning applications.
Key Takeaways
- XENON uses experience-based knowledge correction to enhance planning capabilities.
- The agent integrates an Adaptive Dependency Graph and Failure-aware Action Memory for improved learning.
- XENON outperforms larger proprietary models with a smaller 7B open-weight LLM.
- The approach allows for better handling of flawed priors and sparse feedback.
- Experiments demonstrate significant improvements in knowledge learning and long-horizon planning.
Computer Science > Machine Learning arXiv:2505.24157 (cs) [Submitted on 30 May 2025 (v1), last revised 18 Feb 2026 (this version, v3)] Title:Experience-based Knowledge Correction for Robust Planning in Minecraft Authors:Seungjoon Lee, Suhwan Kim, Minhyeon Oh, Youngsik Yoon, Jungseul Ok View a PDF of the paper titled Experience-based Knowledge Correction for Robust Planning in Minecraft, by Seungjoon Lee and 4 other authors View PDF HTML (experimental) Abstract:Large Language Model (LLM)-based planning has advanced embodied agents in long-horizon environments such as Minecraft, where acquiring latent knowledge of goal (or item) dependencies and feasible actions is critical. However, LLMs often begin with flawed priors and fail to correct them through prompting, even with feedback. We present XENON (eXpErience-based kNOwledge correctioN), an agent that algorithmically revises knowledge from experience, enabling robustness to flawed priors and sparse binary feedback. XENON integrates two mechanisms: Adaptive Dependency Graph, which corrects item dependencies using past successes, and Failure-aware Action Memory, which corrects action knowledge using past failures. Together, these components allow XENON to acquire complex dependencies despite limited guidance. Experiments across multiple Minecraft benchmarks show that XENON outperforms prior agents in both knowledge learning and long-horizon planning. Remarkably, with only a 7B open-weight LLM, XENON surpasses agents that rely...
