Research map / Multi-agent systems and games / LLM agents

Experience-based Knowledge Correction for Robust Planning in Minecraft

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes XENON, a Minecraft planning agent that algorithmically revises its dependency graph and action memory from binary success/failure feedback, letting a 7B LLM outperform agents using much larger proprietary models.

Problem

LLM-based planners in Minecraft start with flawed priors about item dependencies and valid actions, and LLM self-correction from binary failure feedback fails to fix them.

Approach

Instead of asking the LLM to self-correct, the agent algorithmically revises an external dependency graph (Adaptive Dependency Graph, using analogy to successful similar items and inadmissible-item handling) and an action memory that prunes repeatedly failed actions (Failure-aware Action Memory), which also disambiguates dependency errors from action errors.

Key results

• On the MineRL long-horizon benchmark with learned dependencies, XENON (Qwen2.5-VL-7B) reaches 0.54 overall success rate versus 0.45 for Optimus-2 (GPT-4V, oracle dependencies).
• With oracle dependencies, XENON reaches 0.79 overall success rate, versus 0.34 for the Qwen2.5-VL-7B Optimus-1 reproduction.
• In MC-TextWorld under perturbations, XENON keeps EGA near 0.97 while baselines degrade; on TextWorld Cooking XENON reaches 1.00 success rate versus 0.75 for SC and 0.09 for DECKARD.

Surprising findings

• LLM self-correction (SC) achieves the same EGA as DECKARD with no correction in MineRL, plateauing around 0.4, and gives little benefit even in the vanilla MC-TextWorld setting.
• ADAM's EGA is much lower in MineRL (~0.1) than in Mineflayer (~0.6), showing controller capacity strongly affects dependency learning; XENON learns none of the Redstone dependencies in MineRL, yielding 0.00 success on Redstone goals.

Limitations

• Performance depends heavily on the low-level controller; STEVE-1 in MineRL struggles with spatial exploration, creating a gap with Mineflayer.
• Learned dependencies fail for the Redstone group in MineRL because the controller cannot execute the strategy for inadmissible items.
• Several hyperparameters (c0, alpha_i, alpha_s, x0, K) are set by hand and must balance controller robustness against efficiency.

Experimental setup

Datasets: MineRL long-horizon task benchmark (67 goals, 7 groups) MC-TextWorld Mineflayer Microsoft TextWorld Cooking

Benchmarks: MineRL long-horizon task benchmark MC-TextWorld Mineflayer Microsoft TextWorld Cooking

Metrics: Experienced Graph Accuracy (EGA) success rate

Baselines: LLM Self-Correction (SC) DECKARD ADAM RAND DEPS Jarvis-1 Optimus-1 Optimus-2 Optimus-1 (Qwen2.5-VL-7B reproduction)

Builds on: DECKARD Optimus-1 ADAM DEPS

Closest papers by meaning

• ADAM: An Embodied Causal Agent in Open-World Environments (ICLR 2025 · similarity 0.88)
• Experience Transfer for Multimodal LLM Agents in Minecraft Game (CVPR 2026 · similarity 0.87)
• Optimus-1: Hybrid Multimodal Memory Empowered Agents Excel in Long-Horizon Tasks (NeurIPS 2024 · similarity 0.87)
• WALL-E: World Alignment by NeuroSymbolic Learning improves World Model-based LLM Agents (NeurIPS 2025 · similarity 0.87)
• CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI (NeurIPS 2026 · similarity 0.86)
• MrSteve: Instruction-Following Agents in Minecraft with What-Where-When Memory (ICLR 2025 · similarity 0.86)