Why XENON?
1. LLMs often have flawed prior knowledge, resulting in failures to achieve goals (i.e., items in Minecraft).
LLMs often begin with flawed knowledge about dependencies (i.e., requirements) and actions, both of which are essential knowledge for successful planning to achieve goals.
2. LLMs cannot robustly correct their flawed knowledge through simple prompting.
Even when prompted to self-correct with failure feedback, LLMs fail to correct their flawed knowledge.
XENON: Experience-based knowledge correction
XENON corrects its knowledge using its own experience, without additional LLM inference for knowledge correction.
[left] Adaptive Dependency Graph (ADG) is initialized with an LLM's dependency predictions for goal items and revises flawed dependency knowledge. Case 1: If an item is permanently failed to be obtained, it is identified as a hallucinated item and its descendants are recursively revised to remove the flawed dependency. Case 2: A flawed dependency is revised by referencing similar, successfully obtained items.
[right] Failure-aware Action Memory (FAM) corrects action knowledge using past failures, preventing the agent from repeating failed actions.
Dependency and action learning experiments
The y-axis shows the fraction of goal items whose dependencies are correctly learned.
XENON robustly learns knowledge against flawed priors, outperforming previous state-of-the-art agents in dependency and action learning.
Long-horizon goal planning experiments
The numbers in the table show average success rate to achieve goals in MineRL environment.
XENON significantly outperforms previous state-of-the-art agents in long-horizon goal planning.
Citation
@inproceedings{lee2026experience,
title={Experience-based Knowledge Correction for Robust Planning in Minecraft},
author={Lee, Seungjoon and Kim, Suhwan and Oh, Minhyeon and Yoon, Youngsik and Ok, Jungseul},
booktitle={International Conference on Learning Representations},
volume={2026},
pages={137856--137908},
year={2026}
}