01 · High-Information-Density Knowledge Graph Retrieval Augmentation for Large Language Models

针对大型语言模型的高信息密度知识图谱检索增强（LOKG）

Research record	Details
Authors	Zihao Zhou, Ganggang Yin, Zejian Lu, Feng Zhang, Wen Zhao
Published	2026-10-12
Sources	openalex
Focus	knowledge-graph retrieval-augmented generation, token-level information density, schema/triple compression, topology-aware multi-hop retrieval 知识图谱 检索增强生成（RAG）, 令牌级信息密度, 模式/三元组压缩, 拓扑感知多跳检索

English

The paper proposes concrete, implementable modules (schema compression, topology-aware retrieval, triple compression) that address token-budget and efficiency issues directly relevant to building agentic systems with persistent graph memory; its methods and objective (token-level information density) merit careful study for integration and ablation in the thesis experiments.

中文

该文提出了具体可实现的模块（模式压缩、拓扑感知检索、三元组压缩），直接解决与构建具有持久图记忆的智能体系统相关的令牌预算与效率问题；其方法与令牌级信息密度目标值得在论文实验中仔细研究以便集成与消融分析。

Research synopsis

English

The paper diagnoses a practical limitation of retrieval-augmented generation (RAG): attention dilution and token-budget constraints reduce the effectiveness of injecting knowledge into LLM contexts. It proposes LOKG (LLM-Oriented Knowledge Graph), a framework that unifies schema compression, subgraph retrieval, and triple encoding under a token-level information-density objective. LOKG has three modules: SCOS (semantic compression of ontology schema), TAIMR (topology-aware indexing and multi-granularity retrieval), and ITTC (information-theoretic triple compression formulated as a knapsack problem). Experiments on several KGQA benchmarks under a single-LLM-call constraint report improved accuracy–efficiency trade-offs; ablations claim modular orthogonality and synergy. The contribution is an auditable, token-budget-aware KG retrieval pipeline.

中文

论文指出检索增强生成（RAG）在将知识注入大模型上下文时受注意力稀释与令牌预算限制的问题。提出 LOKG（面向 LLM 的知识图谱）框架，将模式压缩、子图检索与三元组编码在令牌级信息密度目标下统一设计。LOKG 包含三部分：SCOS（本体模式的语义压缩）、TAIMR（拓扑感知索引与多粒度检索）和 ITTC（将三元组压缩视作背包问题的信息论方法）。在多套 KGQA 基准、单次 LLM 调用限制下的实验报告了更优的准确性—效率折中；消融研究指出模块间的正交性与协同作用。贡献在于一个可审计、考虑令牌预算的 KG 检索流水线。

Thesis relevance

English

Overlap: both works integrate knowledge graphs into RAG-style pipelines and target multi-hop retrieval efficiency and verifiability. Differences: LOKG emphasizes offline schema compression, token-level information-density optimization, and compressed triple encoding for single-call LLM scenarios; it does not address incremental, persistent graph memory assembled from the system’s own interactions, agentic workflows, or entity-resolution strategies highlighted in the thesis. Complementarity: LOKG’s compression and topology-aware retrieval techniques could be adapted to reduce context cost when serializing subgraphs or reasoning paths from a persistent graph memory, improving end-to-end latency and token-efficiency in the thesis’s agentic experiments.

中文

重合点：两者都将知识图谱整合进 RAG 风格流水线，关注多跳检索的效率与可验证性。差异：LOKG 强调离线的模式压缩、令牌级信息密度优化与在单次 LLM 调用下的三元组压缩编码；它并未处理论文关注的从系统交互增量构建的持久图记忆、智能体工作流或实体解析策略。互补性：LOKG 的压缩方法和拓扑感知检索可用于在将持久图记忆中的子图或推理路径序列化到上下文时降低令牌开销，从而在论文的智能体实验中改善端到端延迟和令牌效率。

Writing and related work

English

Position this paper as an engineering- and information-theory–oriented KG-RAG contribution focused on token-budget trade-offs. Cite it when discussing efficient serialization of graph content, compression-aware retrieval, or single-call LLM constraints.

中文

将该工作定位为以工程与信息论为导向的 KG-RAG 贡献，侧重令牌预算权衡。撰写时在讨论图内容的高效序列化、感知压缩检索或单次 LLM 调用约束时引用该文。

Method and evaluation

English

Consider reusing ITTC’s knapsack formulation to decide how to encode graph triples retrieved from persistent graph memory under a token budget, and adopt TAIMR’s topology-aware path scoring as an alternative ranking function for multi-hop path traversal in the evolving graph. Evaluate how schema compression (SCOS) affects verifiability and bridge-entity identification in multi-hop QA. Measure the impact of these compression modules on latency, retrieval accuracy, and the tendency for graph fragmentation when combined with different entity-resolution strategies.

中文

考虑将 ITTC 的背包式建模复用于在令牌预算下决定如何编码从持久图记忆检索出的三元组，并采用 TAIMR 的拓扑感知路径评分作为在演化图上多跳路径遍历的备选排序函数。评估模式压缩（SCOS）对多跳问答中可验证性和桥接实体识别的影响。度量这些压缩模块与不同实体解析策略结合时对延迟、检索准确率以及图碎片化倾向的影响。

Future directions

English

Integrate LOKG’s modules into an agentic RAG system with an incrementally assembled persistent graph memory and test whether token-aware compression preserves verifiability and multi-hop path reuse over many interactions. Study interactions between compression and entity-resolution errors to understand trade-offs between compactness and graph coherence.

中文

将 LOKG 的模块集成到具有增量构建持久图记忆的智能体 RAG 系统中，验证令牌感知压缩能否在多次交互中保持可验证性和多跳路径重用。研究压缩与实体解析错误之间的相互作用，以理解紧凑性与图连贯性之间的权衡。

---