01 · MOSAIC: Learning Graph Node Embeddings from Spectrally Isolated Dominant Subspaces of Accumulated Diffusion Operators

MOSAIC：从累积扩散算子的谱孤立主子空间学习图节点嵌入

Research record	Details
Authors	Taha Mohseni Ahooyi, Benjamin J. Stear, Yuanchao Zhang, Aditya Lahiri, James Terry, Shiping Zhang, J. Alan Simmons, Ryan Corbett, Patricia Sullivan, Asif Chinwalla, Chris Nemarich, Jo Lynne Rokita, Sharon Diskin, Jonathan C. Silverstein, Deanne M. Taylor
Published	2026-10-04
Sources	openalex
Focus	agent skill for KG querying, biomedical knowledge graph, entity resolution, versioned KG interface 用于KG查询的智能体技能、生物医学知识图谱（KG）、实体解析、可版本化的KG接口

English

Practically useful for engineering reproducible, versioned KG interfaces and query-validation design, but its focus on a curated biomedical KG and packaged skill is less directly relevant to the thesis’s core questions about agent-generated persistent graph memory and entity-resolution trade-offs.

中文

对于工程可复现、可版本化的 KG 接口与查询验证设计具有实用价值，但其侧重于经人工整理的生物医学 KG 和打包技能，与本论文关于由智能体生成的持久图记忆及实体解析权衡这一核心问题的直接相关性较弱。

Research synopsis

English

The paper presents ddkg.skill, an Agent Skill designed to help LLMs query a large, complex biomedical knowledge graph (the NIH Common Fund Data Ecosystem Data Distillery Knowledge Graph, DDKG). ddkg.skill bundles a controller, a compositional library of DDKG-specific references and structured tables, documentation, validated query patterns, a routing table, and a Python checker that tests internal links. Builds are identified by checksum for reproducibility. An orthogonal nine-test evaluation on a live DDKG instance showed partial success (five of seven targeted-source tests executed correctly) and revealed one poorly posed cross-source query plus documentation errors that informed revisions. Three applied biomedical use cases demonstrate the skill’s practical cross-source query workflows.

中文

本文介绍了 ddkg.skill，一种用于辅助大型语言模型查询复杂生物医学知识图谱（NIH Common Fund Data Ecosystem Data Distillery Knowledge Graph，DDKG）的智能体技能。ddkg.skill 将一个控制器、DDKG 专用的组合参考与结构化表格、主文档、经过验证的查询模式、路由表以及用于检查内部链接的 Python 脚本打包在一起。构建通过校验和标识以保证可复现性。在一次针对实时 DDKG 的正交九项测试评估中，针对缺乏示例来源的七项测试中有五项得到正确执行，同时暴露出一次表述不当的跨源比较查询以及技能参考材料中的错误，这些问题推动了修订。文中还给出三个展示跨源查询工作流的生物医学使用案例。

Thesis relevance

English

Overlap: both the thesis and this work address how external knowledge graphs can be made usable by AI systems and address entity-resolution and query construction concerns. Differences: ddkg.skill targets a large, curated biomedical KG and packages static, versioned references and validation logic for LLM-supported querying, whereas the thesis studies an agentic RAG system that incrementally assembles graph memory from the agent’s own retrievals and inferences on local models. Limitation: ddkg.skill does not propose or evaluate persistent, query-accumulating structured graph memory generated by the agent itself, nor compare embedding-based vs. LLM-as-judge entity-resolution trade-offs. Complementarity: the skill’s emphasis on documented, versioned interfaces and query validation could inform reproducibility and verifiability components in the thesis’s framework.

中文

重合点：论文与本论文均关注如何让外部知识图谱（KG）可被 AI 系统有效使用，并触及实体解析与查询构建问题。差异：ddkg.skill 面向大型、经人工整理的生物医学 KG，打包静态的可版本化参考资料与验证逻辑以支持 LLM 查询；而本论文研究的是由智能体自身检索与推理逐步构建的图记忆，并在本地模型上运行的 agentic RAG 系统。局限性：ddkg.skill 并不提出或评估由智能体自身生成并持续累积的长期结构化图记忆，也未比较基于嵌入与 LLM 评判的实体解析权衡。互补性：ddkg.skill 对文档化、可版本化接口与查询验证的重视可以为本论文中关于可验证性和可重复性的设计提供参考。

Writing and related work

English

Position this work as an example of engineering a portable, versioned interface to a complex domain KG that prioritizes reproducible query execution and human-auditable validation. In related work, distinguish between approaches that supply a curated KG interface (ddkg.skill) and those that build persistent graph memory from agent interactions.

中文

将此工作定位为面向复杂领域 KG 的可移植、可版本化接口工程示例，强调可复现的查询执行和可人工审计的验证。在相关工作中，应明确区分提供已整理 KG 接口（ddkg.skill）的方法与通过智能体交互自身构建持久图记忆的方法。

Method and evaluation

English

Consider integrating a ddkg.skill–style controller and routing table into an agentic RAG pipeline to improve query validation and provenance reporting for queries that touch an external KG. Empirically, compare such a validation layer against a baseline without it on the MultiHop-RAG benchmark, and measure effects on bridge-entity identification, verifiability, and end-to-end latency. Also borrow the paper’s checksum/versioning idea to pin graph-memory snapshots for repeatable experiments.

中文

可以将 ddkg.skill 式的控制器与路由表集成到 agentic RAG 管道中，以增强查询验证和针对外部 KG 的溯源报告。在实证评估上，将带有此验证层的系统与无该层的基线在 MultiHop-RAG 基准上比较，并测量其对桥接实体识别、可验证性和端到端延迟的影响。同时借鉴论文的校验和/版本化思路，为图记忆快照上锁以保证可重复实验。

Future directions

English

Adapt the skill concept to support incrementally assembled graph memory: design a versioned packaging and validation layer that can reference agent-constructed subgraphs and record provenance. Evaluate hybrid pipelines that combine ddkg.skill–style validation with embedding- or LLM-based entity-resolution strategies under resource-constrained, local-model settings.

中文

将该技能概念扩展为支持增量组装的图记忆：设计一种可版本化的打包与验证层，使其能够引用由智能体构建的子图并记录溯源。评估在资源受限的本地模型设置下，将 ddkg.skill 式验证与基于嵌入或基于 LLM 的实体解析策略相结合的混合流水线。

---