01 · Towards a structure-aware multi-hop retrieval framework for large language model reasoning over complex oncology guidelines

面向复杂肿瘤指南的结构感知多跳检索框架研究

Research record	Details
Authors	Dongni Li, Guogang Xu, Haitao Huang, Yanjun Zhang, Hui Wang, Lei Wu, Yue Wu, Qingyu Meng, Xian Wang
Published	2026-10-14
Sources	openalex
Focus	structure-aware retrieval, multi-hop retrieval, clinical guideline parsing, structure-aware reranking 结构感知检索, 多跳检索, 临床指南解析, 结构感知重排序

English

The paper presents concrete structure-aware preprocessing and reranking methods that could meaningfully improve triple extraction and noisy-edge reduction for an incrementally assembled graph memory; these techniques and their ablations are directly applicable to the thesis’ ingestion and evaluation design.

中文

该文提出的结构感知预处理与重排序方法可显著改善三元组提取与噪声边减少，直接有助于增量组装图记忆的入库与评估；其方法与消融结果值得深入研读。

Research synopsis

English

The paper introduces SMART-RAG, a structure-aware multi-hop retrieval framework designed for reasoning over National Comprehensive Cancer Network (NCCN) oncology guidelines. It automates structural parsing and performs structured chunking with metadata, then applies structure-aware reranking and multi-hop retrieval to trace evidence across flowcharts, tables, annotations, and cross-references. Evaluations on choice and free-response oncology benchmarks—built from public and manually curated datasets—show improved stability and accuracy relative to conventional RAG, with ablations highlighting the contributions of chunking, reranking, and multi-hop retrieval. The study emphasizes reducing retrieval noise and dynamic cross-page/cross-guideline evidence tracing.

中文

本文提出 SMART-RAG，一种用于对美国国家综合癌症网络（NCCN）指南进行推理的结构感知多跳检索框架。该方法自动进行结构解析并把文档按带元数据的结构化片段化，随后采用结构感知重排序与多跳检索，以追踪流程图、表格、注释与跨页／跨指南的证据。作者在由公开与人工整理数据构成的选择题与自由答题肿瘤学基准上评估，结果较传统的检索增强生成（RAG）展现出更稳定与更高的准确率，消除检索噪声，消融实验表明片段化、重排序与多跳检索的关键作用。

Thesis relevance

English

Overlap: both address multi-hop retrieval and the problem of tracing evidence across complex document structure. Differences: SMART-RAG focuses on automated structural parsing, chunking, and reranking for guideline documents rather than building a persistent, incrementally assembled graph memory; it does not claim agentic interaction, entity-resolution strategies, or long-term structured memory. Complementarity: SMART-RAG’s structural parsing and metadata-enriched chunking could serve as a high-quality document-to-graph ingestion step for the thesis’ graph memory, improving the fidelity of extracted triples and the precision of multi-hop paths.

中文

重合点：两者都关注多跳检索以及在复杂文档结构中追踪证据的问题。差异：SMART-RAG 侧重于对指南文档的自动结构解析、带元数据的片段化与重排序，而非构建持久的、增量装配的图记忆；摘要中未涉及智能体交互、实体解析策略或长期结构化记忆。互补性：SMART-RAG 的结构解析与元数据增强片段化可以作为论文所述图记忆的高质量文档入库步骤，提升提取的三元组与多跳路径的准确性。

Writing and related work

English

Position this paper in related work sections that cover document structure and retrieval: emphasize that structure-aware chunking and reranking mitigate retrieval interference in complex, semi-structured corpora. Cite its ablation approach as an example of isolating preprocessing and retrieval contributions when comparing structured vs. vector retrieval.

中文

在相关工作中将本文列为关注文档结构与检索的方法：强调结构化片段化与重排序在半结构化语料中减轻检索干扰的作用。其消融实验可作为示例，用以在比较结构化检索与向量检索时区分预处理与检索模块的贡献。

Method and evaluation

English

Consider adapting SMART-RAG’s structural parsing and metadata-enriched chunking as an ingestion pipeline for the thesis’ graph memory, converting flowcharts/tables/annotations into candidate triples and provenance links. Evaluate whether structure-aware reranking reduces noisy edges during triple extraction and improves bridge-entity recall on MultiHop-RAG. Report ablations that separate parsing/chunking, reranking, and multi-hop traversal to quantify their individual effects on graph coherence and downstream QA.

中文

可将 SMART-RAG 的结构解析与带元数据的片段化改造为论文图记忆的入库管线，将流程图/表格/注释转为候选三元组与溯源链接。评估结构感知重排序是否能在三元组抽取阶段减少噪声边并提高 MultiHop-RAG 的桥接实体召回。进行将解析/片段化、重排序与多跳遍历拆分开的消融实验，以量化它们对图连贯性与下游问答的独立影响。

Future directions

English

Integrate SMART-RAG’s parser into an agentic pipeline that extracts confidence-tagged triples to seed a persistent graph memory, then measure how structured ingestion affects entity-resolution trade-offs and long-term retrieval improvements. Test on both guideline corpora and the MultiHop-RAG benchmark to assess transfer and domain adaptation.

中文

将 SMART-RAG 的解析器集成入一个智能体管线，提取带置信度标注的三元组以初始化持久化图记忆，进而度量结构化入库对实体解析权衡与长期检索改进的影响。在指南语料与 MultiHop-RAG 基准上测试，以评估迁移能力与领域适配性。

---

02 · Retrieval-Augmented Generation (RAG) Architectures: GraphRAG, Hybrid Reranking, and Context Drift Mitigation

检索增强生成（RAG）架构：GraphRAG、混合重排序与上下文漂移缓解

Research record	Details
Authors	Prakash Prasad, Suraj Kumar, Mohit Mishra, Dr. Vishal Shrivastava, Vibhakar Pathak
Published	2026-10-13
Sources	openalex
Focus	GraphRAG, hybrid retrieval and reranking, context drift, multi-hop retrieval GraphRAG, 混合检索与重排序, 上下文漂移, 多跳检索

English

Highly relevant: provides formalism, hybrid-retrieval recipes, community-detection topology, and empirical benchmarks that can directly inform the thesis’s retrieval baselines, graph-structuring choices, and evaluation metrics.

中文

高度相关：提供的形式化方法、混合检索方案、社区检测拓扑与实证基准可直接为论文的检索基线、图结构选择和评估指标提供借鉴。

Research synopsis

English

The paper analyzes limitations of ‘naive’ flat vector RAG—semantic fragmentation, failure on multi-hop summarization, and context degradation—and presents a formal and empirical study of advanced RAG architectures. It models GraphRAG topology using hierarchical community detection, formulates probabilistic marginalization dynamics of retrieval-generation pipelines, and evaluates hybrid multi-stage retrieval combining dense dual-encoders, sparse lexical search (BM25), Reciprocal Rank Fusion (RRF), and cross-encoder neural rerankers. Experiments on three enterprise corpora (financial 10-Ks, biomedical clinical literature, legal case law) report large gains for GraphRAG and reductions in context hallucination; the paper also highlights open directions including agentic iterative retrieval and provenance verification.

中文

论文分析了“朴素”基于向量的检索增强生成（RAG）的局限性——语义碎片化、多跳全局总结失败及上下文退化，提出对先进 RAG 架构的形式化与实证研究。文中用层次化社区检测刻画 GraphRAG 的拓扑结构，构建了检索—生成流水线的概率边缘化模型，并评估了将稠密双编码器、稀疏词表检索（BM25）、Reciprocal Rank Fusion（RRF）与 cross-encoder 重排序器组合的混合多阶段检索。基于三个企业语料（财务 10-K、医学临床文献、法律判例）的实验显示 GraphRAG 带来显著提升并减少上下文幻觉，论文还提出了智能体迭代检索与溯源验证等开放方向。

Thesis relevance

English

Strong overlap: the paper directly studies GraphRAG and hybrid retrieval strategies relevant to structured graph-based memory and multi-hop retrieval. It contributes a formal probabilistic model and empirical evidence that GraphRAG can substantially improve multi-hop comprehensiveness and reduce hallucination versus flat vector RAG. Limitations relative to the thesis: the abstract emphasizes enterprise document corpora and topological modeling rather than an explicitly incremental, agent-driven graph memory assembled from interaction triples or detailed entity-resolution strategies. The paper complements the thesis by providing retrieval/reranking baselines, formalism for marginalization, and community-detection approaches that could inform graph construction or pruning.

中文

高度重合：该论文直接研究 GraphRAG 与混合检索策略，与基于结构化图记忆和多跳检索的课题高度相关。文中提供了概率形式化模型和实证证据，表明与扁平向量 RAG 相比，GraphRAG 在多跳概括性和减少幻觉方面有显著提升。与论文的差异：摘要侧重企业文档语料和拓扑建模，而不是显式从交互三元组增量构建的智能体驱动图记忆或细化的实体解析方法。该工作可补充本论文的检索/重排序基线、边缘化形式化以及可用于图结构构建或剪枝的社区检测方法。

Writing and related work

English

Cite the paper for formal modeling of retrieval–generation marginalization and for empirical evidence favoring graph-augmented architectures. When positioning your work, contrast their corpus-level GraphRAG topology with your thesis’s incremental, interaction-driven construction and explicit focus on entity-resolution trade-offs.

中文

引用该论文以支持检索—生成边缘化的形式化论述以及实证上对图增强架构的有利证据。在定位你的工作时，应将其基于语料的 GraphRAG 拓扑与本论文侧重的增量交互驱动构建、以及显式的实体解析权衡进行对比。

Method and evaluation

English

Consider adopting their hybrid retrieval pipeline (dense dual-encoders + BM25 + RRF + cross-encoder reranker) as a baseline within your system to isolate the incremental graph-memory contribution. Apply hierarchical community detection ideas to cluster or prune graph regions and to guide multi-hop path search. Incorporate their probabilistic marginalization view to quantify retrieval uncertainty when composing multi-hop reasoning over the evolving KG, and measure metrics analogous to their reported answer comprehensiveness and context-hallucination reductions alongside latency.

中文

可将他们的混合检索流程（稠密双编码器 + BM25 + RRF + cross-encoder 重排序器）作为基线纳入你的系统，以便单独评估增量图记忆的贡献。考虑利用层次化社区检测来对图区域进行聚类或剪枝，并指导多跳路径搜索。将他们的概率边缘化视角引入，以在随时间演进的知识图谱上量化检索不确定性，并在评估中同时记录其所报道的答案完备性、上下文幻觉减少及延迟等度量。

Future directions

English

Extend GraphRAG concepts to an interaction-assembled, persistent graph memory by adapting topology and reranking methods to operate on incrementally added nodes and edges. Experimentally integrate explicit entity-resolution modules (embedding vs. LLM-judge) within the hybrid pipeline and evaluate impacts on graph coherence, multi-hop retrieval, and end-to-end latency.

中文

将 GraphRAG 的思想扩展到基于交互组装的持久图记忆，通过使拓扑与重排序方法适应增量添加的节点与边。将显式实体解析模块（嵌入相似度与 LLM 判定器）整合到混合流程中，实证评估其对图一致性、多跳检索及端到端延迟的影响。

---

03 · BioAgent-RAG: A Knowledge-Graph-Augmented Multi-Agent LLM Framework with Self-Reflective Verification for Explainable Drug Repurposing

BioAgent-RAG：一种知识图谱（KG）增强的多智能体大模型框架，具有自我反思式验证用于可解释的药物再利用

Research record	Details
Authors	Ali Asghari, Abbas Barzegarinezhad
Published	2026-10-13
Sources	openalex
Focus	Knowledge graphs, multi-agent RAG, explainable drug repurposing, factuality verification 知识图谱（KG）, 多智能体 检索增强生成（RAG）, 可解释的药物再利用, 事实性验证

English

Relevant for methods on multi-agent orchestration, path-based KG grounding, and plausibility metrics, but it focuses on curated biomedical KGs rather than the thesis’s incremental, persistent graph memory and entity-resolution trade-offs; useful to skim for methods and evaluation ideas.

中文

对多智能体编排、基于路径的 KG 落地及合理性评估方法具有借鉴价值，但该工作基于人工整理的生物医学 KG，而非论文关注的增量持久化图记忆与实体解析权衡；建议略读以获取方法与评估思路。

Research synopsis

English

This paper addresses hallucination and biological implausibility in biomedical reasoning by proposing BioAgent-RAG, a multi-agent, orchestrated retrieval-augmented generation framework that integrates path-based grounding on existing biomedical Knowledge Graphs (PrimeKG, Hetionet) with autonomous sub-agents for query decomposition and cross-verification (self-reflective verification). The approach emphasizes transparent graph traversal and cross-agent checks to improve plausibility and reduce hallucinations. The authors report empirical gains over baselines (RD2-KG, KGiA, RAG-LLMs) on PrimeKG and Hetionet, providing ROC-AUC, PR-AUC, MRR, Recall@10/50 and domain-specific scores (Biomedical Factuality Score, Path Reasoning Score) and a reported correlation with expert evaluation.

中文

本文针对生物医学推理中的幻觉和生物学不合理论题，提出了 BioAgent-RAG —— 一种多智能体编排的检索增强生成（RAG）框架，结合对现有生物医学知识图谱（PrimeKG、Hetionet）的路径级落地（path-based grounding）、用于查询分解的自治子智能体及交叉验证（自我反思式验证）。该方法强调可解释的图遍历和智能体间的核查以提升合理性并减少幻觉。作者在 PrimeKG 和 Hetionet 上对比 RD2-KG、KGiA、RAG-LLMs 等基线，报告了 ROC-AUC、PR-AUC、MRR、Recall@10/50 以及生物医学事实性评分和路径推理评分，并报告了与专家评估的相关性。

Thesis relevance

English

Overlap: both work on combining Knowledge Graphs with agentic RAG and emphasize transparent, path-based grounding and verifiability. Differences: BioAgent-RAG operates on existing, curated biomedical KGs and targets drug repurposing, whereas the thesis focuses on incrementally assembled graph memory built from the system’s own interactions and persistent graph accumulation. Limitations relative to the thesis: the abstract does not describe persistent, query-aware graph memory construction, nor does it address the entity-resolution trade-offs or consumer-grade local model deployment central to the thesis.

中文

重合点：两者都尝试将知识图谱（KG）与智能体式的检索增强生成结合，强调基于路径的落地与可验证性。差异：BioAgent-RAG 依赖现成的、经人工整理的生物医学知识图谱并以药物再利用为目标；而本论文聚焦于从系统交互中增量构建的持久化图记忆以及图记忆随时间的累积。与论文目标相比的局限：该摘要未描述持久化、查询感知的图记忆构建，也未讨论实体解析的权衡或面向消费者级本地模型部署，这些是论文的核心问题。

Writing and related work

English

Cite this work for methods that combine multi-agent orchestration with KG grounding and for reported domain-specific plausibility metrics. Position it as a domain-focused, explainability-oriented example rather than a direct match for systems that build persistent, incrementally updated graph memory.

中文

可将此工作用于引用多智能体编排与 KG 落地相结合的方法以及领域特定的合理性指标。把它作为一个面向领域、注重可解释性的实例，而非直接匹配于构建持久化、增量更新图记忆的系统。

Method and evaluation

English

Consider adapting their sub-agent decomposition and cross-verification (self-reflective) mechanisms to a setting with incrementally assembled graph memory to evaluate whether verification reduces erroneous edge additions. Reuse their plausibility-oriented metrics (e.g., domain factuality and path reasoning scores) and ranking metrics (ROC-AUC, PR-AUC, MRR, Recall@k) when comparing structured graph memory to vector-only baselines, and measure correlation with expert judgments where available.

中文

可考虑将其子智能体分解与交叉验证（自我反思）机制移植到增量组装的图记忆场景，以评估验证步骤是否能减少错误边的加入。采用其面向合理性的评估指标（例如领域事实性评分和路径推理评分）以及排序指标（ROC-AUC、PR-AUC、MRR、Recall@k）来比较结构化图记忆与仅向量基线，并在可能时测量与专家评估的相关性。

Future directions

English

Follow-ups could apply BioAgent-RAG’s verification and path-grounding techniques to an incrementally constructed, persistent graph memory and explicitly evaluate entity-resolution strategies and runtime/latency trade-offs on consumer-grade local models. Another direction is to generalize the plausibility metrics beyond biomedicine to validate their usefulness in other domains.

中文

后续可将 BioAgent-RAG 的验证与路径落地技术应用到增量构建的持久化图记忆中，并明确评估实体解析策略及在消费者级本地模型上的运行时/延迟权衡。另一方向是将这些合理性指标推广到生物医学以外的领域，以验证其通用性。

---

04 · Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation

单一智能体长期记忆中的 Memory Echo：生命周期评估与基于溯源的缓解

Research record	Details
Authors	Yufeng Li, Sheng Yao, Peng Zeng, Dezhang Kong, Shi Lin, Changting Lin, Xun Wang, Meng Han
Published	2026-10-13
Sources	openalex
Focus	long-term agent memory, Memory Echo, provenance-guided mitigation, benchmark (MEchoBench), decision contamination 长期智能体记忆, 记忆回声（Memory Echo）, 基于溯源的缓解, 基准测试（MEchoBench）, 决策污染

English

Introduces a clearly defined contamination phenomenon (Memory Echo), an evaluation benchmark (MEchoBench), and a mitigation framework (FPA) with quantitative results; these are directly relevant when assessing persistence, provenance, and correction-resistance in long-term agent memory.

中文

提出了明确的污染现象（Memory Echo）、评估基准（MEchoBench）和带量化结果的缓解框架（FPA）；这些内容对评估长期智能体记忆的持久性、溯源性与纠正抗性具有直接参考价值。

Research synopsis

English

The paper defines Memory Echo as excess retrieval exposure caused by repeated storage and retrieval of the same historical claim from a single source, which can bias current agent answers despite lacking independent evidence. To study this, the authors introduce MEchoBench to evaluate echo emergence, correction resistance, persistence, and downstream effects. They propose the Family–Provenance–Authority (FPA) framework: select one representative per source family, require verified corrections before reusing old records, and constrain final answers with authoritative evidence. Paired replay experiments on 190 fixed questions show substantial accuracy gains and reductions in a Decision Contamination Effect metric; code is released.

中文

论文将 Memory Echo 定义为来自同一来源的历史信息经重复存取后获得的过度检索暴露，从而在缺乏独立证据时影响当前智能体回答。为此作者提出 MEchoBench 用以评估回声的产生、抗纠正性、持久性及其对回答的影响。提出 Family–Provenance–Authority（FPA）框架：为每个来源族选择一个代表、在旧记录参与回答前要求已验证的更正、并以权威证据约束最终答案。在对 190 道固定问题的成对重放实验中，FPA 显著提高了准确率并降低了所谓的 Decision Contamination Effect；代码已开源。

Thesis relevance

English

Overlap: both this paper and the thesis investigate risks and mitigation strategies for long-term persistent agent memory and how past records influence current retrieval and answers. Difference: the paper studies same-source accumulation and provenance-based mitigation (Memory Echo) but does not address incremental construction of a structured graph memory, entity-resolution strategies, multi-hop path traversal, or comparison across agentic paradigms and MultiHop-RAG. Complementarity: FPA and MEchoBench offer provenance-aware selection and evaluation artifacts that could be integrated into the thesis’s graph-memory pipeline to reduce contamination and to instrument persistence and correction resistance metrics.

中文

重合点：该论文与本论文均研究长期持久的智能体记忆如何影响当前检索与回答以及可能的缓解策略。差异：该工作聚焦同源记录累积与基于溯源的缓解（Memory Echo），并未涉及增量构建的结构化图记忆、实体消歧方法、多跳路径遍历，或在不同智能体范式与 MultiHop-RAG 上的比较。互补性：FPA 框架与 MEchoBench 提供了面向溯源的选择策略与评估工具，可纳入论文提出的图记忆流水线以降低污染并度量持久性与纠正抗性。

Writing and related work

English

Position this paper in the related work section on persistent agent memory and provenance; cite MEchoBench as an operational benchmark for contamination and correction resistance. Use FPA as an example of provenance-based gating when arguing for mechanisms that constrain reuse of past records in long-lived memory.

中文

在关于持久化智能体记忆与溯源的相关工作中列出此论文；将 MEchoBench 引为评估污染与纠正抗性的实用基准。将 FPA 用作基于溯源的准入示例，以支持在长期记忆中约束历史记录复用的论点。

Method and evaluation

English

Consider adapting FPA’s family selection and correction-conditioning rules to graph-memory maintenance: select or weight a single representative node per source family, and require verified correction flags before reusing historical triples in multi-hop traversal. Reuse the paired-replay evaluation design to measure Decision Contamination Effect and persistence as the graph accumulates, and report how provenance gating affects bridge-entity identification, answer verifiability, and end-to-end latency.

中文

考虑将 FPA 的来源族代表选择与纠正条件规则引入图记忆维护：为每个来源族选择或加权一个代表节点，并在复用历史三元组用于多跳遍历前要求已验证的更正标识。借鉴成对重放的评估设计来测量 Decision Contamination Effect 与随图累积的持久性，并报告溯源准入对桥接实体识别、答案可验证性和端到端延迟的影响。

Future directions

English

Evaluate whether provenance-guided selection (FPA) reduces Memory Echo within an incrementally assembled graph memory, and measure trade-offs with entity-resolution strategies and multi-hop retrieval accuracy on MultiHop-RAG. Extend MEchoBench-style tracks to multi-hop tasks and to local-model deployments used by the thesis.

中文

评估基于溯源的选择（FPA）在增量组装的图记忆中是否能降低 Memory Echo，并测量其与实体消歧策略及在 MultiHop-RAG 上多跳检索准确率的权衡。将 MEchoBench 风格的评测扩展到多跳任务以及论文所用的本地模型部署场景。

---

05 · Text-based Entity Resolution Using Transformer Models

基于Transformer模型的文本实体解析

Research record	Details
Authors	Abrar Ahmed, Li‐Chun Zhang, Ralf Thomas Münnich, Martin Vogt
Published	2026-10-13
Sources	openalex
Focus	entity resolution, transformer models, bi-encoder/cross-encoder cascade, blocking 实体解析, Transformer模型, bi-encoder/cross-encoder级联, 阻塞

English

Highly relevant, practical treatment of the low-latency embedding plus selective refinement strategy that the thesis directly compares; offers implementable techniques and error-assessment methods useful for baselines and evaluation design.

中文

高度相关且具有实践意义，详述了论文将直接比较的低延迟嵌入加选择性精化策略；提供可实现的技术和错误评估方法，对构建基线与设计评估有直接价值。

Research synopsis

English

The paper addresses text-based entity resolution (record linkage) and proposes a practical workflow that combines transformer models with traditional linkage methods to balance accuracy and computational cost. A bi-encoder is used for embedding-based blocking and fast similarity screening; conventional generative or discriminative models perform primary scoring and classification; and a cross-encoder is reserved for difficult candidate pairs to improve final linkage precision. The authors discuss fine-tuning strategies and present an approach for assessing linkage error. Empirical illustrations are given on two benchmark datasets from the Magellan repository.

中文

本文研究基于文本的实体解析（记录链接），提出一个实用的工作流，将Transformer模型与传统链接方法结合以权衡准确性和计算成本。使用bi-encoder进行基于嵌入的阻塞和快速相似度筛选；使用传统的生成式或判别式模型进行主要评分与分类；对于难以判定的候选对则使用cross-encoder以提升最终链接精度。文章讨论了微调策略，并给出一种链接错误评估方法，在Magellan仓库的两个基准数据集上进行了示例性验证。

Thesis relevance

English

Directly relevant to the thesis’s entity-resolution focus: the paper provides a concrete, low-latency embedding-plus-refinement pipeline that matches the thesis’s examined low-latency embedding-similarity strategy. Differences: the paper is framed for static record-linkage datasets and does not address incremental, interaction-driven construction of a knowledge graph, confidence-weighted graph edges, or integration with agentic RAG and multi-hop QA. Complementarity: its cascade design, blocking strategies, and error-assessment techniques can be adapted to the thesis’s graph assembly and evaluation pipeline as a baseline or component of the low-latency branch.

中文

与论文中关于实体解析的研究直接相关：该文章提供了一个具体的低延迟嵌入加精化的工作流，与论文中考察的低延迟嵌入相似性策略相契合。不同点是：该工作面向静态记录链接数据集，未讨论基于交互逐步构建的知识图谱、带置信度的图边或与智能体驱动的RAG及多跳问答的整合。互补之处在于其级联设计、阻塞策略和错误评估技术可被改造并作为论文中低延迟分支的基线或构件。

Writing and related work

English

Cite this paper when discussing practical cascaded designs that trade accuracy for latency; its description of bi-encoder blocking and selective cross-encoder use is a concise justification for adopting a two-stage pipeline. Emphasize that their evaluation is on standard record-linkage benchmarks rather than on downstream multi-hop QA or incremental graph coherence.

中文

在论述为权衡准确性与延迟而设计的实用级联架构时可引用此文；其对bi-encoder阻塞和选择性使用cross-encoder的描述可为采用两阶段流水线提供简洁依据。需要强调的是，他们的评估基于常规模式的记录链接基准，而非下游多跳问答或增量图一致性的场景。

Method and evaluation

English

Adapt the paper’s bi-encoder for fast blocking to generate candidate node-merge proposals and use embedding thresholds to assign preliminary edge confidence scores. Reserve a stronger model (cross-encoder or an LLM judge) for ambiguous merges and calibrate its outputs to produce weighted edges. Evaluate effects on graph fragmentation (number of distinct node clusters per ground-truth entity), downstream multi-hop retrieval accuracy, and end-to-end latency under a cascaded policy.

中文

将论文中的bi-encoder用于快速阻塞以生成候选节点合并提议，并使用嵌入阈值为初步图边分配置信度。对于模糊合并保留更强的模型（cross-encoder或LLM判定器），并对其输出进行校准以产生带权重的边。用图碎片化指标（每个真实实体对应的不同节点簇数量）、下游多跳检索准确率以及在级联策略下的端到端延迟来评估影响。

Future directions

English

Extend the cascade to an online, incremental setting where bi-encoder blocking runs continuously on newly extracted triples and a higher-precision verifier (cross-encoder or LLM judge) is invoked conditionally. Investigate calibration of confidence scores so they can be interpreted as edge weights in an evolving knowledge graph and measured for their impact on multi-hop path integrity.

中文

将该级联方法扩展到在线增量场景：对新抽取的三元组持续运行bi-encoder阻塞，并有条件地调用更高精度的验证器（cross-encoder或LLM判定器）。研究置信度分数的校准，使其可作为演化知识图谱中的边权，并测量其对多跳路径完整性的影响。

---