01 · EngramGraph: Summary-Indexed Sub-Graph Paging for Model-Agnostic Multi-Agent Memory

EngramGraph：面向模型不可知多智能体记忆的基于摘要索引的子图分页

Research record	Details
Authors	Shubham Kunwar Tiwary
Published	2026-10-08
Sources	openalex
Focus	persistent graph memory, summary-indexed sub-graph paging, model-agnostic multi-agent memory, heterogeneous-LLM handoffs 持久化图记忆, 基于摘要索引的子图分页, 模型不可知多智能体记忆, 异构LLM切换

English

The paper proposes a concrete, system-level architecture (summary-indexed sub-graph paging) with strong empirical claims and available code; a deep read is needed to assess implementation details, partitioning algorithms, summary generation, state-continuity mechanisms, and how applicable it is to the thesis’ triple-extraction and entity-resolution needs.

中文

该文提出了具体的系统级架构（基于摘要的子图分页）并给出显著的实证性主张且附有代码；需要深入阅读以评估实现细节、子图划分算法、摘要生成、状态连续性机制，以及其在论文所关注的三元组抽取和实体解析场景中的适用性。

Research synopsis

English

The paper introduces EngramGraph, a model-agnostic persistent graph memory architecture that mitigates context-window saturation and loss of structural memory in long-horizon multi-agent systems. It partitions a shared property graph into semantic Sub-Topic sub-graphs and keeps only two-sentence Sub-Topic summaries in active memory; agents explicitly page-in a full sub-graph on demand, commit typed graph mutations and updated summaries, then page-out. EngramGraph claims substantial active token reduction (66.1%–87.7%), improved multi-hop causal recall (+34.9% vs Top-K vector RAG), and 90%–100% state continuity across heterogeneous LLM handoffs. Implementation references FalkorDB and the Model Context Protocol (MCP).

中文

论文提出 EngramGraph，一种模型不可知的持久化图记忆架构，旨在缓解长时限多智能体系统中的上下文窗口饱和和结构化记忆丢失问题。它将共享属性图划分为语义子主题子图，仅在活动记忆中保留两句子主题摘要；智能体按需显式分页（page-in）完整子图，提交类型化图变更并更新摘要，完成后再分页（page-out）。论文声称显著减少活动提示令牌（66.1%–87.7%）、提高多跳因果召回（比 Top-K vector RAG 高 +34.9%），并在异构 LLM 切换间保持 90%–100% 的状态连续性。实现涉及 FalkorDB 与 Model Context Protocol (MCP)。

Thesis relevance

English

Overlap: EngramGraph addresses persistent graph memory and long-term state continuity, directly relevant to the thesis question of structured graph memory for multi-hop retrieval. Differences: EngramGraph emphasizes sub-graph paging, summary indexes, and model-agnostic heterogeneous-LLM handoffs, while the thesis focuses on incrementally assembled Knowledge Graphs from triple extraction, entity resolution strategies, and comparisons between agentic RAG paradigms. Limitations relative to the thesis: the abstract does not describe triple extraction, entity-resolution methods, or evaluation on a QA benchmark like MultiHop-RAG, so its applicability to entity-fragmentation and bridge-entity reuse is unclear.

中文

重合点：EngramGraph 关注持久化图记忆和长期状态连续性，这与论文关于用于多跳检索的结构化图记忆的研究问题直接相关。差异：EngramGraph 强调子图分页、摘要索引和模型不可知的异构 LLM 切换，而本论文侧重于从交互中增量构建的知识图谱（以三元组为单位）、实体解析策略及不同智能体 RAG 范式的比较。相对论文的局限：摘要未说明三元组抽取、实体解析方法或在像 MultiHop-RAG 之类的 QA 基准上的评估，因此其对实体碎片化及桥接实体复用的适用性尚不清楚。

Writing and related work

English

Position EngramGraph as a systems-level memory-management approach for persistent graph stores: its summary-indexed paging is a useful contrast to always-loading whole graphs or Top-K text fragments. When citing, highlight its model-agnostic design and claimed empirical gains, but note that details on semantic extraction and entity-resolution are not present in the abstract.

中文

可将 EngramGraph 置于系统级的记忆管理方法中：其基于摘要的分页策略可作为与始终加载完整图或 Top-K 文本片段的有益对照。引用时应强调其模型不可知的设计和宣称的实证收益，但需指出摘要中未提供语义抽取与实体解析的具体细节。

Method and evaluation

English

Consider adopting EngramGraph’s summary-indexed sub-graph paging as a memory-management layer for the thesis’s persistent graph memory to reduce prompt token usage and manage latency. Evaluate how paging affects multi-hop retrieval accuracy and reasoning-step reuse on the thesis’ MultiHop-RAG benchmark, and instrument state continuity metrics when switching local models (e.g., Ollama-hosted Qwen2.5). Also test whether sub-graph boundaries interact with entity-resolution strategies and graph fragmentation.

中文

可考虑将 EngramGraph 的基于摘要的子图分页作为论文持久化图记忆的内存管理层，用以降低提示令牌消耗并控制延迟。在论文使用的 MultiHop-RAG 基准上评估分页对多跳检索准确度和推理步骤复用的影响，并在切换本地模型（例如 Ollama 上的 Qwen2.5）时记录状态连续性指标。还应测试子图边界是否与实体解析策略和图碎片化发生交互影响。

Future directions

English

Combine EngramGraph-style paging with the thesis’s incremental triple extraction and explicit entity-resolution experiments to study how paging policies affect graph coherence and bridge-entity reuse. Investigate automated Sub-Topic partitioning tuned for multi-hop QA and measure trade-offs between summary fidelity, page-in latency, and end-to-end QA accuracy.

中文

将 EngramGraph 式的分页与论文中的增量三元组抽取和实体解析实验结合，研究分页策略如何影响图的一致性和桥接实体复用。探索为多跳问答优化的自动子主题划分，并度量摘要保真度、分页延迟与端到端 QA 准确率之间的权衡。

---

02 · A GraphRAG-Enhanced Multi-Agent Framework for Blast Design in Drill-And-Blast Tunnels

用于掘进爆破设计的 GraphRAG 强化多智能体框架

Research record	Details
Authors	李 庆刚, Xuewei Li, Xiaochuan Han, Zhu Dapeng
Published	2026-10-09
Sources	openalex
Focus	GraphRAG, multi-agent systems, multimodal domain adaptation, provenance-aware plan generation GraphRAG、多智能体系统、多模态领域自适应、带出处的计划生成

English

Relevant for practical techniques (LoRA adaptation, multimodal fusion, provenance-rich structured outputs) and comparative GraphRAG vs conventional RAG results, but the paper is an application-specific engineering study that does not address the thesis’s core questions about persistent incremental graph memory or entity-resolution trade-offs.

中文

对实用技术（LoRA 适配、多模态融合、带出处的结构化输出）以及 GraphRAG 与常规模式 RAG 的对比结果有参考价值，但该论文属于面向应用的工程研究，并未涉及论文的核心问题：持久的增量图记忆或实体消歧权衡。

Research synopsis

English

The paper presents a domain-specific multi-agent framework that integrates multimodal LLMs with graph-based retrieval-augmented generation (GraphRAG) to connect geological assessment and preliminary blast-plan generation for drill-and-blast tunnelling. A geological-analysis agent is LoRA-fine-tuned on 1,222 cycles to classify rock mass from images, GPR profiles, and logs; a blast-design agent retrieves code provisions, rules, and cases from a domain knowledge graph to produce structured plans with source and applicability records. Reported results show improved classification and plan macro-F1 compared to unadapted and conventional-RAG baselines, supported by ablations and an evidence-budgeted evaluation, and include an expert evaluation of an illustrative plan.

中文

该论文提出了一个面向隧道掘进爆破设计的多智能体框架，将多模态大模型与基于图的检索增强生成（GraphRAG）结合，用以衔接地质评估与初步爆破方案生成。地质分析智能体通过 LoRA 在 1,222 个循环上进行微调，从隧道面图像、GPR 剖面和地质记录中对围岩分类；爆破设计智能体从领域知识图谱检索规范、设计规则与历史案例，生成带出处与适用性记录的结构化方案。论文在与未适配模型及常规模式 RAG 的对比中报告了分类和方案 macro-F1 的提升，并通过消融实验与有限证据预算评估进行了支持，还包含对实例方案的专家评分。

Thesis relevance

English

Overlap: both use GraphRAG-style retrieval, multi-agent decomposition, and explicit provenance for generated outputs. The paper demonstrates domain adaptation and multimodal fusion that could inform agent perception modules. Differences/limitations: this work targets a curated domain KG and supervised adaptation for a particular engineering task; it does not claim to build or evaluate an incrementally assembled, persistent graph memory accumulated across user interactions, nor to study entity-resolution trade-offs or multiple agentic paradigms. Complementarity: its LoRA adaptation, provenance-rich structured outputs, evidence-budget evaluation, and ablation methodology are directly applicable to the thesis’s experimental design.

中文

重合点：二者都采用 GraphRAG 风格的检索、多智能体分工以及对生成结果的明确出处记录。该论文展示的领域自适应与多模态融合可为论文中智能体的感知模块提供参考。差异/局限：该工作面向已构建的领域知识图谱并通过监督微调解决特定工程任务；它并未声明构建或评估那种通过交互逐步累积的持久图记忆，也未研究实体消歧的权衡或比较多种智能体范式。互补性：其 LoRA 适配、带出处的结构化输出、证据预算评估与消融方法对论文的实验设计具有直接借鉴价值。

Writing and related work

English

Position this paper as an application-style validation of GraphRAG and multimodal domain adaptation rather than as evidence for incremental graph memory or entity-resolution techniques. Emphasize its reporting of provenance and ablation when contrasting against purely vector RAG baselines.

中文

将该论文定位为对 GraphRAG 与多模态领域自适应的应用验证，而非关于增量图记忆或实体消歧方法的直接证据。在与纯向量 RAG 基线对比时，强调其出处记录与消融分析的呈现。

Method and evaluation

English

Adoptable methods include LoRA-based supervised adaptation for perception agents, multimodal input fusion, and producing structured outputs that pair decisions with source and applicability metadata to improve verifiability. Evaluation techniques to borrow: macro-F1 for classification/plan tasks, controlled ablations over input modalities, and an evidence-budget constraint when comparing GraphRAG vs conventional RAG. Do not assume the paper’s KG is incrementally constructed—treat its retrieval component as a retrieval oracle for testing structured output and provenance ideas.

中文

可借鉴的方法包括用于感知型智能体的 LoRA 监督适配、多模态输入融合，以及生成将决策与出处和适用性元数据配对的结构化输出以提升可核查性。可借用的评估手段：用于分类/方案任务的 macro-F1、对输入模态的受控消融实验，以及在比较 GraphRAG 与常规模式 RAG 时施加的证据预算约束。注意不要将论文中的知识图谱视为增量构建——应将其检索组件视为用于检验结构化输出与出处机制的检索资源。

Future directions

English

Extend this engineered GraphRAG setup toward an incrementally assembled graph memory by instrumenting the retrieval component to emit triples and provenance during interaction. Evaluate entity-resolution strategies and latency trade-offs (embedding similarity vs LLM judgment) within the same pipeline, and test agentic paradigms (e.g., ReAct, Self-Ask) for composing multimodal observations into multi-hop retrievals.

中文

将该工程化的 GraphRAG 方案扩展为可增量构建的图记忆，方法是让检索组件在交互中导出三元组与出处。在线管道内评估实体消歧策略与延迟权衡（嵌入相似度 vs LLM 裁判），并测试不同的智能体范式（如 ReAct、Self-Ask）如何将多模态观测组织为多跳检索。

---

03 · PARM: Passage-Anchored Relation-Aware Multi-Hop Retrieval for Evidence-Grounded Question Answering

PARM：基于段落锚点的关系感知多跳检索用于证据驱动问答

Research record	Details
Authors	Muzhi Wang, Fajie Wu, Guangyue Jia, Lei Han, Ruohan Shi, Haozheng Zhu, Zhaobo Qi, Feng Xu
Published	2026-10-09
Sources	openalex
Focus	multi-hop retrieval, passage-anchored graph retrieval, relation-aware path ranking, evidence provenance 多跳检索, 段落锚点图检索, 关系感知路径排序, 证据可溯源

English

PARM presents concrete methods for creating complementary graph anchors and for joint path ranking with provenance; these techniques are directly usable to improve the thesis’s retrieval and path-scoring modules and deserve careful examination.

中文

PARM 提供了生成互补图锚点和基于溯源的联合路径排序的具体方法；这些技术可直接用于改进论文的检索与路径评分模块，值得仔细研读。

Research synopsis

English

PARM addresses multi-hop question answering by introducing passage-level evidence anchors before graph reasoning. The method first retrieves relevant passages via lexical retrieval and semantic reranking, then augments question entities with entities linked to those passages to form complementary graph search anchors. A bounded multi-hop graph search generates explicit relation paths under hop and path budgets; those paths are ranked by semantic relevance, question–relation compatibility, and graph structural features. The system assembles selected passages, relation paths, and provenance into a traceable evidence package and reports empirical improvements on MuSiQue and subsets of 2WikiMultiHopQA.

中文

PARM 通过在图推理前引入段落级证据锚点来解决多跳问答问题。该方法先通过词汇检索和语义重排找出相关段落，然后将问题实体与这些段落关联的实体结合，构成互补的图搜索锚点。受限的多跳图搜索在预定义的跳数与路径预算下生成明确的关系路径；随后基于语义相关性、问题-关系兼容性和图结构特征对路径进行排序。最终选出的段落、关系路径及其溯源被打包为可追溯的证据，且在 MuSiQue 和 2WikiMultiHopQA 子集上报告了实验提升。

Thesis relevance

English

PARM overlaps with the thesis on using explicit relation paths and provenance-aware graph search for multi-hop retrieval. It differs because PARM constructs passage-anchored search seeds per query and evaluates bounded graph search, rather than assembling a persistent, incrementally updated graph memory across interactions. PARM does not address entity-resolution strategies for maintaining long-term graph coherence or agentic orchestration; however, its passage-anchoring and path-ranking signals could be integrated into the thesis’s graph-memory retrieval and path-scoring components.

中文

PARM 在使用显式关系路径与可溯源的图搜索以完成多跳检索方面与论文有重叠。不同之处在于 PARM 为每个查询构建基于段落的搜索种子并评估受限图搜索，而非在交互间增量构建并保持持久的图记忆。PARM 并未处理用于维护长期图一致性的实体解析策略或智能体编排；但其段落锚点生成和路径排序信号可被整合进论文提出的图记忆检索与路径评分模块。

Writing and related work

English

Position PARM among graph-based retrieval works that focus on seed selection and path-ranking rather than persistent knowledge bases. Emphasize its explicit evidence-package output as a useful provenance baseline when comparing verifiability of graph memory versus vector-only RAG.

中文

将 PARM 放在以检索种子选择与路径排序为重点的图检索工作中，而非持久知识库研究。强调其显式证据包输出，作为在比较图记忆与纯向量 RAG 可验证性时的有益基线。

Method and evaluation

English

Consider adopting PARM’s passage-anchoring step to initialize or augment incremental graph memory when answering a new query, and reuse its joint ranking features (semantic relevance, question–relation compatibility, graph-structural scores) to score candidate paths in the persistent graph. Evaluate bounded hop/path budgets as ablation variables in the thesis’s MultiHop-RAG benchmark and measure effects on bridge-entity identification and end-to-end latency.

中文

可考虑采用 PARM 的段落锚点步骤来初始化或增强增量图记忆以响应新查询，并复用其联合排序特征（语义相关性、问题-关系兼容性、图结构评分）对持久图中的候选路径进行评分。在论文使用的 MultiHop-RAG 基准上将受限跳数/路径预算作为消融变量进行评估，并测量其对桥接实体识别和端到端延迟的影响。

Future directions

English

Adapt PARM’s passage-anchoring and path-ranking techniques to an incremental, persistent graph memory setting and measure whether these signals help reduce graph fragmentation. Extend its provenance packaging to include node/edge confidence scores and entity-resolution traces used by the thesis’s graph memory.

中文

将 PARM 的段落锚点与路径排序技术适配到增量持久图记忆场景中，并评估这些信号是否有助于减少图片段化。将其证据打包扩展为包含节点/边置信度以及论文图记忆中使用的实体解析轨迹。

---

04 · Pricing the Context Window: A Unified Write–Read Theory for Token-Efficient Agent Memory

为上下文窗口定价：一种用于令牌高效智能体记忆的统一写-读理论

Research record	Details
Authors	Frank W. Bergmann
Published	2026-10-09
Sources	openalex
Focus	agent memory optimization, write–read policies, context-token pricing, information-theoretic memory 智能体记忆优化, 写-读策略, 上下文令牌定价, 信息论记忆

English

Provides a formal, actionable framework for token-aware write/read trade-offs and a depth law directly relevant to planning traversal depth and budgeting in graph-based agentic RAG systems; despite lack of empirical validation, its propositions can shape experimental design and hypotheses in the thesis.

中文

提供了关于令牌感知写/读权衡的形式化框架和与遍历深度规划直接相关的深度定律；尽管缺乏实证验证，其命题可用于指导论文的实验设计与假设检验。

Research synopsis

English

The paper frames the finite context window of an LLM-based agent as a scarce resource and formalizes agent memory as a joint write–read investment problem. It couples write policies (what and at which resolution to store) and read policies (what to load for a query) by pricing context tokens with a shadow price λ. Building on known ingredients (value of information, submodular selection, rate–distortion, MDL), the contribution is the coupled treatment and its formal implications: a depth law for optimal traversal depth, a write criterion comparing storage to re-derivation cost, a lower bound on information content for predictively sufficient memory, and thresholds for resolution, caching, and indexing. Results are presented as propositions and eight testable hypotheses but are not empirically validated.

中文

论文将基于大模型的智能体的有限上下文窗口视为稀缺资源，并将智能体记忆形式化为联合的写-读投资问题。作者以影子价格 λ 对上下文令牌定价，耦合写策略（存储什么、以何种分辨率）与读策略（为查询加载什么）。在已知成分（信息价值、次模选择、率失真、MDL）基础上，主要贡献是耦合分析及其形式含义：最优遍历深度的“深度定律”、将写入与重推导成本比较的写入判据、可预测充分记忆信息含量的下界，以及对分辨率、缓存和索引的阈值条件。结果以命题和八个可检验假设呈现，但未做实证验证。

Thesis relevance

English

Overlap: both address optimization of agent memory and make explicit the trade-off between what to store and what to re-derive, and the paper’s depth law directly speaks to query planning and traversal depth in multi-hop graph retrieval. Differences/limitations: the paper is theoretical and token-centric, offering no empirical validation and not addressing structured, incrementally assembled graph memory, entity resolution, or KG-specific fidelity. Complementarity: the theory can inform the thesis’s write/read policies, token-budget-aware traversal depth, and evaluation of latency vs. accuracy trade-offs when accumulating graph memory.

中文

重合点：二者都探讨智能体记忆的优化，并明确写入与重推导之间的权衡；论文中的深度定律与查询规划及多跳图检索的遍历深度直接相关。差异/局限：该工作以理论为主、以令牌为中心，缺乏实证验证，且并未涉及结构化、增量构建的图记忆、实体解析或知识图谱（KG）特有的一致性问题。互补性：其理论框架可为论文中的写/读策略、基于令牌预算的遍历深度决策及在积累图记忆时对延迟与准确性权衡的评估提供指导。

Writing and related work

English

Position this paper as a formal, economics-flavored framework for memory budgeting in the related-work section. Note its emphasis on token pricing and planning rather than on structured KG construction or entity resolution; this distinguishes it from empirical graph-memory work.

中文

在相关工作部分将该论文定位为关于记忆预算的形式化、带有经济学风格的框架。指出其侧重于令牌定价与规划而非结构化知识图谱构建或实体解析，以此将其与实证的图记忆工作区分开。

Method and evaluation

English

Translate the depth law into experiments that vary a shadow price λ (or effective token budget) and measure resulting traversal depth, retrieval accuracy, and latency on MultiHop-RAG. Implement write-criteria that compare stored triple value vs. simulated re-derivation cost and measure graph growth, fragmentation (entity-resolution effects), and end-to-end QA performance. Use the paper’s thresholds as hypotheses to test rather than assumed design rules.

中文

将深度定律转换为实验：在 MultiHop-RAG 上改变影子价格 λ（或有效令牌预算），测量由此产生的遍历深度、检索准确率和延迟。实现将存储的三元组价值与模拟的重推导成本比较的写入判据，评估图的增长、碎片化（实体解析影响）及端到端问答性能。把论文中的阈值作为待检验的假设，而非既定设计规则。

Future directions

English

Empirically validate the propositions in the context of persistent query-aware graph memory by mapping token costs to actual latency and model-compute budgets. Extend the model to incorporate entity-resolution costs and confidence-weighted edges, and test whether the predicted thresholds hold when memory is structured as an incrementally built KG.

中文

在持久的查询感知图记忆场景中，将令牌成本映射到实际延迟与计算预算，以实证验证论文的命题。将模型扩展为包含实体解析成本与置信加权边，并测试当记忆以增量构建的知识图谱（KG）形式存在时，论文预测的阈值是否成立。

---

05 · From Retrieval to Evidence: A Task-Adaptive Two-Layer Knowledge Architecture for LLM Agents

从检索到证据：面向任务的两层知识架构用于大型语言模型智能体

Research record	Details
Authors	Aleksandr Volkov, Taras Pustovoy, Dmitriy Istomin, Tatiana Otbetkina
Published	2026-10-09
Sources	openalex
Focus	two-layer knowledge architecture, evidence admission, task-adaptive routing, retrieval vs structured memory 两层知识架构, 证据准入, 任务自适应路由, 检索与结构化记忆比较

English

The paper presents a tested architecture and empirical findings about task-conditional trade-offs between compact reviewed knowledge and versioned evidence admission—topics directly relevant to grounding and evidence management in graph-memory research. Its methods and production-scale observations are likely to yield actionable ideas for the thesis.

中文

该文提出了经过测试的架构并给出关于精简经审核知识体与版本化证据准入之间任务条件性权衡的实证发现——这些主题与图记忆研究中的落地与证据管理直接相关。其方法与生产级观测很可能为论文提供可操作的思路。

Research synopsis

English

The paper studies grounding strategies for LLM agents, arguing that retrieved similarity and graph association each have complementary strengths that vary by task. It proposes a task-adaptive two-layer architecture: a compact, reviewed Body of Knowledge (BoK) for orientation and synthesis, and a versioned evidence layer for exact lookup and claim support. A router selects which representation to use and an independent admission stage validates named entities, time, source role, contradiction, and provenance. Controlled studies across several course-like benchmarks and a large observational case are reported, showing task-conditional performance trade-offs and substantial reduction in standing representation size.

中文

该论文研究用于大型语言模型智能体的落地（grounding）策略，指出检索相似性与图关联在不同任务中各有优势且互补。作者提出一套面向任务的两层知识架构：用于定向与跨源合成的精简、经审核的知识体（Body of Knowledge, BoK），以及用于精确查证与支撑主张的版本化证据层。一个路由器决定使用何种表示，独立的准入阶段校验命名实体、时间、来源角色、矛盾与可溯源性。文中在若干课程式对照实验和一个大规模观测案例上报告了任务条件下的性能权衡，并展示了常驻表示的大幅缩减。

Thesis relevance

English

Directly relevant: both works compare structured memory and retrieval-based grounding and study when each is advantageous. Complementary: the paper emphasizes a two-layer BoK+versioned-evidence design and an explicit evidence admission gate, which could be integrated with an incrementally assembled graph memory. Differences/limitations: the paper does not emphasize incremental graph construction, entity-resolution strategies, multi-hop path traversal, or comparisons among agentic paradigms—central concerns of the thesis.

中文

直接相关：两者都比较了结构化记忆与基于检索的落地策略，并研究了各自在何种任务中更有利。互补之处：本文强调 BoK+版本化证据的两层设计以及独立的证据准入门控，这一思路可以与增量组装的图记忆结合。差异/局限：本文并不突出增量图构建、实体消歧策略、多跳路径遍历或不同智能体范式间的比较——而这些是本论文的核心关切。

Writing and related work

English

Emphasize the paper’s separation of candidate discovery, packet selection, and evidence admission when positioning related work; cite its task-conditional crossover finding to argue that representation choice should be evaluated per-task. Use its production-scale observational example to motivate scalability and gating design.

中文

在撰写相关工作时突出该文将候选发现、包选择与证据准入分离的做法；引用其任务条件下的交叉发现以支持“按任务评估表示选择”的论点。可利用其生产级观测示例来论证可扩展性与门控设计的必要性。

Method and evaluation

English

Consider incorporating a router+admission gating mechanism into the thesis system to decide when to consult query-aware graph memory versus BoK or raw RAG results. Evaluate this hybrid policy on the thesis’s MultiHop-RAG benchmark, measuring bridge-entity identification, end-to-end latency, and how admission gates affect precision/recall. Use the versioned evidence layer idea to record provenance and allow rollbacks or confidence-based selection of graph edges.

中文

可考虑将路由器+准入门控机制并入论文系统，用于决定何时查询查询感知图记忆、BoK或原始 RAG 结果。在论文使用的 MultiHop-RAG 基准上评估该混合策略，度量桥接实体识别、端到端延迟，以及准入门控对精确率/召回率的影响。借鉴版本化证据层的思路记录可溯源信息，以便回滚或基于置信度选择图边。

Future directions

English

Integrate the two-layer BoK/evidence gating with an incrementally constructed graph memory to test whether admission checks improve entity resolution and multi-hop path coherence. Explore adapting the router to be agent-aware so different agentic paradigms (e.g., ReAct, Self-Ask) can trigger different grounding strategies.

中文

将两层 BoK/证据门控与增量构建的图记忆结合，检验准入检查是否能改善实体消歧和多跳路径连贯性。探索使路由器具备智能体感知能力，从而允许不同智能体范式（例如 ReAct、Self-Ask）触发不同的落地策略。

---

06 · An Evidence-Based Knowledge Graph Framework with Large Language Models and Human-in-the-Loop for Ovarian Cancer Resistance

一种结合大型语言模型与人类参与的基于证据的卵巢癌耐药知识图谱框架

Research record	Details
Authors	Horacio Thompson, Rocío Ayelem Conforti, Sebastián A. Andújar, Marilina Casais, Carlos Marcelo Telleria, Marcelo Luis Errecalde
Published	2026-10-09
Sources	openalex
Focus	evidence-grounded knowledge graphs, LLM–KG integration, human-in-the-loop, biomedical GraphRAG and QA 基于证据的知识图谱（KG）, 大型语言模型与KG集成, 人类参与（Human-in-the-Loop）, 生物医学 GraphRAG 与问答

English

The paper offers concrete methods for evidence linking, expert-driven coherence evaluation (Gwet’s AC1), and GraphRAG-style QA adjudication that are directly useful for designing verifiability, provenance, and evaluation in a thesis about graph memory and agentic RAG.

中文

该论文提供了用于证据链接、专家驱动一致性评估（Gwet’s AC1）和 GraphRAG 式问答判定的具体方法，对构建图记忆与智能体 RAG 的可验证性、来源记录和评估设计具有直接参考价值。

Research synopsis

English

The paper proposes an Evidence-based Knowledge Graph (E-KG) framework that integrates large language models, Knowledge Graphs (KGs), and a Human-in-the-Loop scheme to construct, evaluate, and reason over triples explicitly linked to supporting textual evidence. The authors apply the framework to ovarian cancer resistance: they used MedGemma-27B for extraction, involved experts throughout the E-KG lifecycle with interactive visualization tools, and evaluated coherence and biological plausibility using an expert-driven protocol (Gwet’s AC1). They complement expert evaluation with GraphRAG-style tasks (link prediction metrics and a question-answering case study judged by both LLM-as-a-judge and expert audit) and report producing a coherent, auditable E-KG specialized for this domain.

中文

本文提出了证据型知识图谱（E-KG）框架，将大型语言模型、知识图谱（KG）与人类参与相结合，用以构建、评估并基于证据文本链接进行推理。作者在卵巢癌耐药方向应用该框架：使用 MedGemma-27B 提取三元组，贯穿生命周期引入专家并配套交互式可视化工具，采用专家驱动协议（Gwet’s AC1）评估一致性与生物学合理性。论文还通过 GraphRAG 风格任务（链路预测指标与一个问答案例）并由 LLM-as-a-judge 与专家审计共同评判，构建出面向该领域的可审计 E-KG。

Thesis relevance

English

Overlap: both works combine LLMs and KGs and evaluate GraphRAG-style tasks with attention to provenance and verifiability. Differences/limitations: the paper is domain-specialized (ovarian cancer) and emphasizes expert-in-the-loop curation and evidence linking rather than agentic, incremental graph memory or comparisons among agentic paradigms and entity-resolution strategies. Complementarity: the paper’s methods for explicit evidence linking, expert-driven coherence metrics (Gwet’s AC1), and LLM-plus-expert QA adjudication can inform the thesis’s verifiability, provenance bookkeeping, and evaluation design.

中文

重合点：两者都将大型语言模型与知识图谱结合，并通过 GraphRAG 式任务关注可追溯性与可验证性。差异/局限：该论文为领域专用（卵巢癌），侧重专家参与的人工审校与证据链接，而非智能体驱动的增量图记忆、对比多种智能体范式或专门比较实体解析策略。互补性：其关于显式证据链接、专家驱动一致性度量（Gwet’s AC1）以及 LLM 与专家共审的问答判定方法，可为论文在可验证性、来源记录和评估设计上提供参考。

Writing and related work

English

Position this paper as an example of strong provenance and human-auditing practice when arguing for verifiability in graph memory systems. Cite its use of expert-driven coherence metrics and explicit triplet-to-text linking as methodological precedents, while noting the domain-specific scope.

中文

在论述图记忆系统的可验证性时，可将该论文作为良好来源和人工审计实践的示例。引用其专家驱动一致性度量和三元组到文本的显式链接方法，但同时指出其领域专用性。

Method and evaluation

English

Consider adopting explicit textual provenance for each extracted triple as in this work to improve auditability. Use expert-driven coherence assessment with inter-annotator agreement measures (e.g., Gwet’s AC1) alongside automated GraphRAG evaluations (Hits@k, MRR) and an LLM-as-judge plus expert audit workflow to assess faithfulness and relevance. MedGemma-27B is an example LLM for extraction; report the model used and the human-review protocol when comparing entity-resolution strategies.

中文

可借鉴本文为每个提取三元组保存显式文本来源以增强可审计性。将专家驱动的一致性评估（含注释者一致性度量，如 Gwet’s AC1）与自动化 GraphRAG 评测（Hits@k、MRR）以及 LLM-as-a-judge 加专家审计的流程并行使用，以评估忠实性与相关性。MedGemma-27B 为提取任务的示例模型；在比较实体解析策略时应报告所用模型与人工审阅协议。

Future directions

English

Extend the evidence-grounded E-KG approach to non-biomedical domains or to agentic, incrementally assembled graph memory to test scalability and cross-domain coherence. Combine the human-in-the-loop provenance pipeline with automated entity-resolution techniques and measure their trade-offs in graph coherence, retrieval accuracy, and latency on multi-hop benchmarks.

中文

将该证据型 E-KG 方法推广到非生物医学领域或集成到智能体驱动的增量图记忆中，以测试可扩展性和跨域一致性。把人类参与的证据管线与自动化实体解析方法结合起来，在多跳基准上测量图一致性、检索准确率与延迟之间的权衡。

---