From entity-centric to goal-oriented graphs: Enhancing LLM knowledge retrieval in minecraft

Jonathan Leung, Yongjie Wang, Zhiqi Shen

TL;DR

This paper introduces Goal-Oriented Graphs (GoGs), a novel knowledge representation for Large Language Models (LLMs) that structures procedural knowledge around goals and their dependencies. By moving beyond fragmented entity-centric graphs, GoGs enable LLMs to perform robust multi-step reasoning, significantly outperforming existing methods like GraphRAG in complex tasks like Minecraft.

What Problem Does This Paper Address?

Large Language Models (LLMs) excel at many tasks but struggle with step-by-step procedural reasoning, especially in complex, interactive environments. Existing retrieval-augmented generation (RAG) methods, such as GraphRAG, attempt to address this by structuring knowledge into entity-relation graphs. However, these entity-centric approaches often fragment causal knowledge into numerous low-granularity triples, making it difficult to reconstruct coherent, multi-step plans. This fragmentation hinders an agent's performance in domains requiring intricate planning, like Minecraft.

Methods

The core contribution is the Goal-Oriented Graph (GoG) framework. Unlike traditional entity-centric graphs, GoGs represent goals as nodes and logical dependencies (subgoal relationships) as directed edges. This structure explicitly captures procedural dependencies.

The GoG construction involves:

  1. Goal Extraction: Using LLMs to extract goals, their attributes (description, preconditions, postconditions), and subgoal relationships from unstructured text sources (e.g., Minecraft Wiki).
  2. Goal Merging and Subgoal Derivation: Merging duplicate goals and establishing complete subgoal relationships across different text chunks. This involves comparing name and condition similarities between extracted goals.

During inference, a goal-driven retrieval algorithm traverses the GoG. Given a task query, it retrieves relevant goals, recursively identifies subgoals, and constructs a coherent reasoning chain to guide the LLM's plan generation. This hybrid approach combines symbolic reasoning (GoG structure) with LLM-based semantic grounding.

Datasets

The primary dataset used for training and evaluation is the Minecraft environment and tasks. Specifically, the authors utilize tasks provided by Li et al. (2024), comprising 66 tasks categorized into 7 groups (wood, stone, iron, gold, diamond, redstone, and armor), ordered by increasing difficulty. The knowledge base for GoG and GraphRAG is constructed from Minecraft Wiki pages and in-game recipe files.

Key Insights and Ideas

  • Goals over Entities for Procedural Reasoning: The paper strongly argues that for procedural tasks, representing knowledge as goals and their dependencies (GoGs) is far more effective than representing it as fragmented entities and relations (GraphRAG). This is because procedural tasks are inherently goal-driven.
  • Coherent Reasoning Chains: GoGs enable the explicit retrieval of causal reasoning paths by recursively breaking down high-level goals into subgoals. This creates a structured, coherent chain that guides the LLM, preventing the "shredded paper" problem of fragmented knowledge.
  • Robustness to Imperfect Knowledge: While GoGs perform best with clean, semi-structured data, they demonstrate graceful degradation when faced with noisy or incomplete procedural knowledge, maintaining relatively high soundness and completeness.
  • LLM as Both Extractor and Translator: The LLM plays a dual role: extracting structured goal information from unstructured text and then translating abstract GoG dependencies into executable action descriptions for the agent.

Key Results

  • Significant Performance Improvement: GoG substantially improves procedural reasoning and task success in Minecraft, significantly outperforming GraphRAG and other baselines, especially in more complex task groups (iron, gold, armor, diamond).
  • Success Rate: For challenging tasks like gold and diamond, GoG maintained a success rate above 50%, while baselines consistently failed within the maximum allowed steps. For the gold task, GoG's success rate was three times higher than HKG.
  • Efficiency: GoG generally achieved higher success rates and lower required steps compared to baselines, indicating more efficient planning.
  • GraphRAG Pitfalls: GraphRAG sometimes performed worse than the Vanilla baseline, likely due to retrieving irrelevant but connected nodes, leading to noisy context that degraded task execution.
  • Plan Quality: Ablation studies showed that including both the goal tree and material/tool lists significantly improved plan quality (goal satisfaction, soundness, completeness, and efficiency) across all LLMs.

Limitations and Future Work

  • Reliance on Extractable Procedural Knowledge: GoG's effectiveness depends on the availability of procedural knowledge that can be structured into goal-precondition-effect formats. In domains where this knowledge is absent, highly implicit, or non-symbolic, its benefits may be reduced.
  • Generalization to Unfamiliar Domains: While tested in Minecraft, assessing GoG's performance in domains unfamiliar to pretrained LLMs is an important future direction.
  • Multiple Reasoning Paths: The current implementation retrieves a single coherent reasoning chain. Future work could explore quantitative analysis of path diversity, alternative goal chains, and optimality-aware retrieval (e.g., cost- or success-weighted edges).
  • Online Replanning and Scalability: The study focused on offline graph construction. Future work will investigate incremental online replanning and scalability strategies (goal clustering, semantic de-duplication, graph sparsification) for large-scale or cross-domain settings.
  • Creative/Abstract Goals: GoG could be extended to handle creative or abstract goals by leveraging LLMs' generative capabilities to dynamically construct and refine GoGs for novel tasks.

Summary updated

Abstract

• A Goal-Oriented Graph (GoG) framework for LLM procedural reasoning. • Models goal dependencies to overcome fragmentation of entity-centric graphs. • Goal-driven retrieval constructs coherent reasoning chains to guide LLM agents. • Significantly outperforms state-of-the-art GraphRAG in complex Minecraft tasks. Large Language Models (LLMs) demonstrate impressive general capabilities but often struggle with step-by-step procedural reasoning, a critical challenge in complex interactive environments. While retrieval-augmented methods like GraphRAG attempt to bridge this gap, their fragmented entity-relation graphs hinder the construction of coherent, multi-step plans. In this paper, we propose a novel framework based on Goal-Oriented Graphs (GoGs), where each node represents a goal and edges encode logical dependencies between them. This structure enables the explicit retrieval of causal reasoning paths by identifying a high-level goal and recursively retrieving its prerequisites, forming a coherent chain to guide the LLM. Through extensive experiments on the Minecraft testbed, a domain that demands robust multi-step planning and provides rich procedural knowledge, we demonstrate that GoG substantially improves procedural reasoning and significantly outperforms GraphRAG and other state-of-the-art baselines.
BibTeX
@inproceedings{leung2026from,
  author = {Jonathan Leung and Yongjie Wang and Zhiqi Shen},
  title = {{From entity-centric to goal-oriented graphs: Enhancing LLM knowledge retrieval in minecraft}},
  booktitle = {Knowledge-Based},
  year = {2026},
  doi = {10.1016/j.knosys.2026.115706},
}
Figure