01 · Persistent State and Becoming. When Memory Becomes Operational History

持久状态与“成为”：当记忆成为操作性历史

Research record	Details
Authors	Zawwar Sami
Published	2026-10-07
Sources	openalex
Focus	long-term agent memory, operational history, representation-adequacy test, relational ablations 长期智能体记忆, 操作性历史, 表征充分性测试, 关系消融

English

The paper provides a focused conceptual framework and concrete tests (representation-adequacy and counterfactual pre/post-state) that are directly applicable to evaluating whether accumulated graph memory changes agent behavior; these ideas are likely to inform evaluation design and ablation strategies in the thesis.

中文

该文提出了针对评估保留记忆是否改变智能体行为的概念框架与具体测试（表征充分性与反事实前/后状态），可直接用于设计论文的评估与消融策略，因而值得深入阅读。

Research synopsis

English

The paper examines when retained records constitute an operational past that materially changes a deployed system’s present behavior. It distinguishes five layers of retained information—archive, context, persistent state, operational history, and becoming—and argues that records only become operational history when relations among events (order, supersession, verified outcome, provenance) change what the system should do now. The author proposes two formal tests: a representation-adequacy test and a counterfactual pre/post-state test that requires durable, attributable behavioral change (without model-weight updates). The essay situates this account among long-term memory work and specifies a controlled M1–M4 relational-ablation evaluation; no benchmark results are reported.

中文

本文探讨何时保留的记录构成能够实质性改变部署系统当前行为的操作性过去。作者区分了五个信息层次——档案、语境、持久状态、操作性历史与“成为”，并主张仅当事件之间的关系（顺序、取代、验证结果、溯源）改变了系统现在应当执行的行为时，记录才成为操作性历史。文中提出两种形式化测试：表征充分性测试与反事实前/后状态测试，后者要求可持久、可归因的行为改变（不依赖模型权重更新）。文章将该观点置于长期智能体记忆研究语境中，并给出受控的 M1–M4 关系消融评估；未报告基准结果。

Thesis relevance

English

Overlap: both address long-term agent memory and ask when retained information should influence future behavior, which is central to evaluating persistent graph memory. Differences: this paper is primarily conceptual and evaluative, offering formal tests and ablation schemas rather than implementing a graph-memory system or reporting empirical benchmarks. Complementarity: its representation-adequacy and counterfactual tests can be operationalized to evaluate whether the thesis’s incrementally built graph memory causes durable, attributable changes in retrieval and QA behavior.

中文

重合点：两者都关注长期智能体记忆以及何时保留信息应影响后续行为，这对评估持久图记忆至关重要。差异：该文主要为概念性与评估性工作，提供形式化测试与消融方案，但未实现图记忆系统或报告实证基准。互补性：其表征充分性和反事实测试可以被用来检验论文中增量构建的图记忆是否会在检索与问答行为上产生持久且可归因的改变。

Writing and related work

English

Cite this paper when motivating evaluation criteria for persistent graph memory and when distinguishing record survival from functional adaptation. Use its layered vocabulary (archive → becoming) to clarify claims about what kinds of stored triples should affect agent decisions.

中文

在论证持久图记忆的评估标准或区分记录保留与功能性适应时引用此文。可采用其分层词汇（档案→“成为”）来澄清哪些存储的三元组应当影响智能体决策。

Method and evaluation

English

Incorporate the paper’s representation-adequacy test and counterfactual pre/post-state test into empirical evaluation: run paired experiments where the system’s graph memory is ablated (M1–M4 style) and measure whether behavior on MultiHop-RAG queries changes in a durable, attributable way without retraining. Instrument provenance, supersession, and event ordering in the graph and report cases where retained relations alter retrieval paths or final answers.

中文

将论文的表征充分性测试与反事实前/后状态测试纳入实证评估：进行配对实验，对系统的图记忆按 M1–M4 风格消融，并测量在不重训练的情况下，MultiHop-RAG 查询的行为是否发生持久且可归因的改变。为图中记溯源、取代与事件顺序建立记录，并报告保留的关系如何改变检索路径或最终答案的示例。

Future directions

English

Apply the proposed tests empirically to the thesis’s persistent graph memory, designing automated metrics for ‘‘operational history’’ detection and extending M1–M4 ablations to entity-resolution components. Investigate remediation policies for when retained graph relations produce incorrect operational changes.

中文

将所提测试实证应用于论文的持久图记忆，设计用于“操作性历史”检测的自动化指标，并将 M1–M4 消融扩展到实体解析组件。研究在保留图关系导致错误操作性改变时的补救策略。

---

02 · A Survey on Retrieval-Augmented and Transformer-Based Approaches for Legal AI: Grounding, Reasoning, and Multilingual Outcome Prediction

关于法律人工智能中检索增强与基于 Transformer 方法的综述：语义接地、推理与多语种结果预测

Research record	Details
Authors	Neeraja R A, Sherin Shaji, Devende M R, Ajmalsha S R, Mrs. Viji C
Published	2026-10-07
Sources	openalex
Focus	retrieval-augmented generation (RAG), Transformer embeddings, legal question answering, multilingual outcome prediction 检索增强生成（RAG）, Transformer 嵌入, 法律问答, 多语种结果预测

English

The survey gives useful, focused critiques of grounding, evaluation, and dataset limitations that are practically relevant, but its narrow scope (four core studies) and domain focus on legal AI make it lower priority than system papers directly addressing graph memory and agentic RAG.

中文

该综述对接地、评估与数据集局限给出有价值的聚焦性批判，对实验设计有参考价值，但其仅基于四项核心研究且面向法律领域，优先级低于直接讨论图记忆与智能体 RAG 的系统性论文。

Research synopsis

English

This paper is a focused technical survey of retrieval-augmented and Transformer-based approaches applied to legal AI tasks such as legal question answering and multilingual outcome prediction. The authors synthesize four recent peer-reviewed studies (LQRAG, LegalReasoner, an embedding-LLM pairing evaluation on Indian statutes, and a stacked-ensemble multilingual predictor), critically assessing methodologies, sample sizes, metrics, and limitations. They report scope-limited evidence that retrieval grounding, domain-adapted embeddings, and ensembles can improve within-study results, and identify recurring gaps: small/proof-of-concept datasets, reliance on proprietary LLM evaluators, lack of domain-expert human evaluation, narrow language/jurisdiction coverage, and temporal drift in legal knowledge. The review is explicitly not a systematic literature review and bounds its conclusions accordingly.

中文

本文对应用于法律 AI 的检索增强与基于 Transformer 的方法进行了聚焦性技术综述，涵盖法律问答与多语种结果预测等任务。作者综合并批判性评估了四项近期经审稿研究（LQRAG、LegalReasoner、在印度法规上的嵌入–LLM 配对评估，以及一个多语种堆叠集成预测器）的研究方法、样本规模、评价指标与局限性。综述发现有限证据表明检索接地、领域适配的嵌入和集成方法在各自研究中可提升表现，但存在共同缺陷：数据集规模小或概念验证性质、依赖专有 LLM 作为自动评估器、缺乏法律领域专家的人类评估、语言与司法管辖覆盖狭窄，以及法律知识的时序漂移。作者声明该综述并非系统性文献综述，其结论因此受限。

Thesis relevance

English

Overlap: the survey discusses RAG, grounding, multi-stage reasoning, and empirical evaluation choices that relate to retrieval and verifiability concerns in the thesis. Differences: it is a domain-focused survey in law rather than an empirical systems paper about agentic RAG with persistent structured graph memory, and it does not address graph construction, entity resolution, or agentic paradigms (ReAct, Plan-and-Execute, Self-Ask, Reflexion). Limitations: the review is based on four core studies with bounded scope and does not evaluate structured long-term memory or multi-hop graph traversal. Complementarity: its critical points on grounding, evaluator selection, and dataset limitations can inform experimental design and evaluation for the thesis’s systems and benchmarks.

中文

重合点：该综述讨论了 RAG、接地策略、多阶段推理与实证评估选择，這些主题与论文中关于检索和可核验性的关注相关。差异：该工作是面向法律领域的综述，而非关于具有持久结构化图记忆的智能体 RAG 的实证系统研究，并且未讨论图构建、实体解析或智能体范式（ReAct、Plan-and-Execute、Self-Ask、Reflexion）。局限性：综述以四项核心研究为基础，范围受限，未评估结构化长期记忆或多跳图遍历。互补性：其对接地、评估器选择和数据集局限的批判性洞见可用于改进论文的实验设计与评价。

Writing and related work

English

Positioning lesson: clearly state selection criteria and scope when presenting a focused survey to avoid overgeneralization. Emphasize transparent reporting of dataset sizes, evaluation metrics, and evaluator provenance (human vs. proprietary automated).

中文

写作提示：在做聚焦性综述时应明确选择标准与范围以避免过度泛化。强调透明报告数据规模、评价指标与评估者来源（人工评估与专有自动评估）的重要性。

Method and evaluation

English

For empirical work, compare retrieval grounding strategies across embedding–LLM pairings and include human expert adjudication where possible. When evaluating structured graph memory, measure multi-hop retrieval accuracy, bridge-entity identification, latency, and verifiability against vector-only and non-persistent agentic baselines; borrow the review’s caution about small datasets and report per-language/jurisdiction breakdowns.

中文

在实证研究中，比较不同嵌入–LLM 配对下的检索接地策略，并尽量纳入领域专家的人工裁决。评估结构化图记忆时，应将多跳检索准确率、桥接实体识别、延迟与可核验性与纯向量基线和无持久记忆的智能体基线进行对比；参考综述对小规模数据集的警示，并按语言/司法管辖区分解报告结果。

Future directions

English

Extend the survey scope to more datasets and jurisdictions, add legal-domain human evaluation, and study how persistent structured graph memory and agentic RAG perform in legal tasks—particularly with respect to verifiability, temporal drift, and multilingual embedding adaptation.

中文

未来工作可将综述扩展至更多数据集与司法管辖区、加入法律领域的人类评估，并研究持久结构化图记忆与智能体 RAG 在法律任务中的表现，重点关注可核验性、时序漂移与多语种嵌入适配。

---

03 · Generative Artificial Intelligence-Driven Enterprise Knowledge Base

生成式人工智能驱动的企业知识库

Research record	Details
Authors	邹志新, Jiahui Zheng, Xiyin Zheng, Portia Cobbinah, Rogozhina Mariia, Shishan Quan, Bojing Liu
Published	2026-10-07
Sources	openalex
Focus	knowledge graph embeddings, generative question answering, reinforcement-learning feedback, enterprise knowledge base 知识图谱（KG）嵌入, 生成式问答, 基于强化学习的反馈, 企业知识库

English

Relevant for KG-embedding and RL-optimized generation components but lacks the thesis’s core focus on agentic, persistent graph memory, entity-resolution trade-offs, and local model/agentic paradigm comparisons; useful to skim for components but not central for deep-read.

中文

对知识图谱嵌入和强化学习优化的生成组件具有参考价值，但缺乏论文的核心关注点：智能体的持久图记忆、实体消歧权衡以及本地模型/智能体范式比较；适合略读以获取可借用的组件思路，而非深读。

Research synopsis

English

The paper proposes KQGen, a unified framework for enterprise knowledge bases that combines knowledge graph embeddings with generative question answering. KQGen encodes KG entities and relations, performs query understanding, and produces answers grounded in structured knowledge; its generation component is optimized with a reinforcement learning-based feedback mechanism. The authors report improvements on relevance and ranking metrics on benchmarks including SQuAD, TREC, and KBQA, while noting remaining challenges in Exact Match (EM) performance and scalability. Future work plans include multi-hop reasoning refinement and multi-modal integration.

中文

论文提出了 KQGen，一个将知识图谱（KG）嵌入与生成式问答相结合的企业知识库框架。KQGen 对实体与关系进行编码，执行查询理解，并基于结构化知识生成答案；生成模块通过基于强化学习的反馈机制进行优化。作者在 SQuAD、TREC 和 KBQA 等基准上报告了在相关性和排序指标上的提升，同时指出 Exact Match（EM）性能和可扩展性仍存在挑战。后续工作拟改进多跳推理并整合多模态数据。

Thesis relevance

English

Overlap: both works use knowledge-graph representations and generative QA methods and care about grounding answers in structured knowledge. Differences: KQGen focuses on enterprise KG embeddings and RL-optimized generation rather than agentic workflows, persistent or incrementally constructed graph memory, or explicit entity-resolution strategies. Complementarity: KQGen’s embedding and RL-feedback techniques could be adapted as components for node/edge encoding or answer grounding within the thesis’s agentic graph-memory architecture, but the paper does not address multi-session graph accumulation, bridge-entity reuse, or comparisons among agentic paradigms.

中文

重合点：两者都使用知识图谱表征与生成式问答方法，并关注基于结构化知识的答案支撑。差异：KQGen 侧重企业场景下的 KG 嵌入与强化学习优化的生成，而不是智能体工作流、持久或增量构建的图记忆，也未针对显式的实体消歧策略进行探讨。互补性：KQGen 的嵌入与 RL 反馈方法可作为本论文图节点/边编码或答案支撑的组件，但该论文未涉及多会话图累积、桥接实体重用或智能体范式比较。

Writing and related work

English

Position KQGen as representative of embedding+generative QA approaches for enterprise KBs; cite it when discussing prior work that grounds LLM output in KG encodings and uses RL for generation tuning. Note that such papers often emphasize relevance/ranking but do not treat persistent agentic memory or explicit entity-resolution trade-offs.

中文

将 KQGen 作为企业知识库中嵌入+生成式问答方法的代表；在论述将 LLM 输出与 KG 编码结合并用强化学习优化生成的前沿工作时可引用。需指出此类工作常侧重相关性/排序，而不关注持久的智能体记忆或明确的实体消歧权衡。

Method and evaluation

English

Consider reusing KQGen-style KG embeddings to initialize node/edge feature vectors and to provide query-aware candidate scoring in the graph memory. Evaluate whether RL-based feedback for generation can be adapted to produce confidence signals used as edge weights. In empirical comparisons, measure not only ranking and relevance (as KQGen does) but also bridge-entity identification, graph coherence, multi-hop QA accuracy, and end-to-end latency on MultiHop-RAG or equivalent multi-document, bridge-entity annotated benchmarks.

中文

可考虑复用 KQGen 风格的 KG 嵌入以初始化图中节点/边的特征向量，并作为查询感知的候选评分信号。评估是否能将基于强化学习的生成反馈改造成可作为边权重的置信度信号。在实证比较中，应不仅衡量相关性/排序（KQGen 的关注点），还应在 MultiHop-RAG 或同类含桥接实体标注的多文档基准上测量桥接实体识别、图一致性、多跳问答准确率和端到端延迟。

Future directions

English

Follow-ups could adapt KQGen’s embedding and RL-feedback components to an incrementally constructed graph memory, using generation feedback to calibrate edge confidence; extend the framework to explicit entity-resolution modules and evaluate effects on multi-hop path recovery and latency. Exploring deployment with local models and privacy-preserving constraints would also align it with the thesis scope.

中文

后续工作可将 KQGen 的嵌入与强化学习反馈组件整合到增量构建的图记忆中，利用生成反馈来校准边的置信度；将框架扩展为包含显式的实体消歧模块，并评估其对多跳路径恢复和延迟的影响。将系统在本地模型和隐私保护约束下部署亦可使其与论文范围对齐。

---

04 · TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning

TopoGraphRAG-Bench：评估基于版面布局的多模态 GraphRAG 在证据推理上的表现

Research record	Details
Authors	Ruochi Li, Jianzhe Lin, Haoxuan Zhang, Haihua Chen, Junhua Ding, Edward Gehringer, Yang Zhang
Published	2026-10-07
Sources	openalex
Focus	GraphRAG, multimodal document reasoning, evidence topology, layout-grounded retrieval, topology-aware evaluation GraphRAG, 多模态文档推理, 证据拓扑, 版面布局定锚检索, 拓扑感知评估

English

The benchmark, controlled topologies, counterfactual validation, and topology-aware metrics are highly relevant and potentially reusable for evaluating whether an incrementally assembled graph memory preserves multi-hop evidence paths—material that could materially affect experimental design and evaluation in the thesis.

中文

该基准、受控拓扑、反事实验证与拓扑感知指标高度相关且有可复用性，可用于评估增量组装的图记忆是否保留多跳证据路径——这些内容可能显著影响论文的实验设计与评估方案。

Research synopsis

English

The paper introduces TOPOGRAPHRAG-BENCH, a layout-grounded benchmark for multimodal GraphRAG that targets recovery of evidence topology across text, tables, figures and captions in long, visually rich documents. It contains 2,024 questions over 201 documents, with questions constructed under three controlled topologies: single-hop retrieval, bridge-chain reasoning, and multi-source synthesis; counterfactual validation is applied to ensure modality and evidence necessity. The authors evaluate text-only GraphRAG, page-level visual retrieval, and multimodal GraphRAG using retrieval, generation, and topology-aware reasoning metrics. Results show multimodal GraphRAG outperforms baselines but still fails when visual–text alignment or unit composition is incomplete, motivating models that explicitly model layouts and cross-modal alignment.

中文

本文提出 TOPOGRAPHRAG-BENCH，一种基于版面布局的多模态 GraphRAG 基准，聚焦于在长篇视觉丰富文档中恢复跨文本、表格、图表与图注的证据拓扑。基准包含 201 篇文档上的 2,024 个问题，问题按单跳检索、桥链推理和多源合成三类受控拓扑构造，并对问题施加反事实验证以保证模态与证据的必要性。作者比较了文本-only GraphRAG、页级视觉检索与多模态 GraphRAG，使用检索、生成及拓扑感知推理指标评价。结果显示多模态 GraphRAG 整体最强，但在视觉-文本对齐或多单元组合不完整时仍失效，强调需要显式建模版面与跨模态对齐。

Thesis relevance

English

Overlap: both work on GraphRAG-style architectures and on recovering multi-hop/bridge-chain evidence topologies. The benchmark’s controlled topologies and topology-aware metrics (and its counterfactual validation) are directly applicable for evaluating whether a persistent structured graph memory preserves intended reasoning paths. Differences: this paper focuses on multimodal, layout-grounded documents and offline benchmark evaluation, not on agentic systems, incremental assembly of graph memory from interactive agent activity, entity-resolution strategies, or comparisons among agent paradigms on local LLMs. Complementarity: layout and cross-modal alignment techniques and topology metrics here can be incorporated into the thesis’s graph memory representation and evaluation suite.

中文

重合点：两者都涉及 GraphRAG 式架构和多跳/桥链证据拓扑的恢复。该基准的受控拓扑、拓扑感知评估指标与反事实验证可直接用于评估持久化结构化图记忆是否保留预期的推理路径。差异：该论文侧重于多模态、基于版面的文档与离线基准评估，而不涉及智能体驱动的系统、由交互生成的增量图记忆、实体消歧策略，或在本地 LLM 上比较不同智能体范式。互补性：其版面建模、跨模态对齐方法和拓扑指标可被纳入论文拟议的图记忆表示与评估工具集中。

Writing and related work

English

Position this paper in related-work as evidence that GraphRAG evaluations should measure topology recovery and modality necessity, not just passage relevance or answer quality. When arguing for structured graph memory, cite their failure modes (visual–text alignment and multi-unit composition) to motivate richer node/edge types and layout-aware alignment.

中文

在相关工作中把该论文列为证据，说明 GraphRAG 的评估应衡量拓扑恢复与模态必要性，而不仅是段落相关性或答案质量。在论述结构化图记忆的必要性时，可引用其失败模式（视觉-文本对齐与多单元组合）来支持引入更丰富的节点/边类型与版面感知的对齐方案。

Method and evaluation

English

Adopt the paper’s controlled topologies (single-hop, bridge-chain, multi-source) and its counterfactual validation when designing QA evaluation to test whether accumulated graph memory preserves bridge entities and paths. Treat visual/table/figure units as distinct node types in the graph and evaluate cross-modal alignment failure modes; use topology-aware metrics (beyond QA F1) to quantify recovery of intended reasoning paths.

中文

在设计 QA 评估时采用该文的受控拓扑（单跳、桥链、多源）与反事实验证，以测试累积的图记忆是否保留桥实体与路径。将视觉/表格/图像单元视为图中的不同节点类型并评估跨模态对齐的失效模式；使用拓扑感知指标（超出 QA F1）来量化对预期推理路径的恢复能力。

Future directions

English

Extend the thesis’s persistent graph memory to include layout-grounded, multimodal evidence units and evaluate whether incremental accumulation improves topology recovery. Explore visual–text co-reference resolution by incorporating layout and visual features into entity-resolution pipelines and node embeddings.

中文

将论文中的持久化图记忆扩展为包含版面定锚的多模态证据单元，并评估增量累积是否提升拓扑恢复能力。通过将版面和视觉特征纳入实体消歧流程与节点嵌入，探索视觉-文本共指的解析方法。

---

05 · The Homeostatic Substrate Hypothesis Persistent Agent Memory as Ecological Tissue

稳态基质假说：作为生态组织的持久性智能体记忆

Research record	Details
Authors	John R. Smith, SHAI / HATI
Published	2026-10-09
Sources	openalex
Focus	persistent agent memory, homeostatic modelling of memory dynamics, memory entrenchment and revision, measurement (Memory Drift Index, MDI) 持久性智能体记忆, 记忆动力学的稳态建模, 记忆固化与修正, 度量（记忆漂移指数 MDI）

English

Recommend a skim: the paper offers a useful conceptual framing (HSH), a prescriptive measurement instrument (MDI), and engineering motifs (LoopGuard) that inform evaluation and safety design for persistent memory, but it lacks concrete KG construction or retrieval methods directly applicable to the thesis.

中文

建议略读：该论文提供了有用的概念框架（HSH）、可采纳的度量工具（MDI）和工程化思路（LoopGuard），可丰富对持久性记忆的评估与安全设计，但缺乏可直接用于本论文的知识图谱构建或检索方法。

Research synopsis

English

The paper reports a conceptual and measurement-oriented proposal motivated by experimental observations that agents with persistent, human-readable text memory show dynamics analogous to physiological/ecological homeostasis failures: accumulation, entrenchment of false beliefs, deletion-sensitive function, and occasional self-correction. The authors propose the Homeostatic Substrate Hypothesis (HSH), framing persistent agent memory as regulatory “tissue” whose coherence and failure modes should be modelled rather than treated as a neutral store. They introduce the Memory Drift Index (MDI) as a pre-registered, exposure-corrected metric for confidence-tagged memory invalidation, outline formal foundations drawn from prior programme documents, state pre-committed hypotheses with kill conditions, and argue for engineering consequences (an immune-like LoopGuard). The paper emphasizes that HSH is a testable framing, not an established mechanism.

中文

该论文基于实验观察提出了一个概念性与度量导向的方案：具有持久、人类可读文本记忆的智能体表现出类生理/生态稳态失衡的动力学特征——记忆累积、错误信念的固化、对删除敏感的功能以及偶发的自我纠正。作者提出稳态基质假说（HSH），将持久性智能体记忆视为具有调节功能的“组织”，应对其相干性与失效模式进行建模而非当作中性存储。文中引入记忆漂移指数（MDI）——一个针对带置信标签记忆失效的、经曝光校正的预注册度量，说明了来自既有方案文件的形式基础，给出预先提交的假设与终止条件，并主张工程后果（如免疫式机制 LoopGuard）。作者强调 HSH 是待检验的表述，而非已证实机理。

Thesis relevance

English

Overlap: both works concern persistent agent memory and risks of accumulated incorrect beliefs that impair downstream reasoning. Differences: this paper is primarily a high-level hypothesis and measurement proposal about memory dynamics (HSH, MDI) rather than a methods paper about constructing or querying structured graph memory (知识图谱（KG）) or entity-resolution techniques. Complementarity: MDI and the immunological LoopGuard idea could be adapted to evaluate and mitigate entrenchment of false triples or relations in an incrementally assembled graph memory; the paper provides framing and pre-registered evaluation discipline that could strengthen the thesis’s long-term-memory evaluation.

中文

重合点：两者都关注持久性智能体记忆及其累积错误信念对后续推理能力的影响。差异：该论文主要提出有关记忆动力学的高阶假说与度量（HSH、MDI），而不是关于构建或查询结构化图记忆（知识图谱（KG））或实体解析方法的技术性论文。互补性：MDI 与免疫式 LoopGuard 概念可用于评估并缓解逐步构建的图记忆中错误三元组或关系的固化；论文提供的理论框架与预注册评估规范可增强本论文对长期记忆稳定性和回归风险的考察。

Writing and related work

English

Position this work in related work as a conceptual contribution on long-term memory dynamics and failure modes. Cite HSH/MDI when motivating safeguards, deletion/validation policies, and pre-registered measurement in the thesis; note that the paper does not supply KG-specific construction or traversal algorithms.

中文

在相关工作中将此作为关于长期记忆动力学与失效模式的概念性贡献。撰写时在论证保障措施、删除/验证策略和预注册度量时引用 HSH/MDI；并指出该论文并未提供针对知识图谱（KG）构建或遍历的具体算法。

Method and evaluation

English

Adopt MDI-like metrics to quantify accumulation of low-confidence or invalid triples in graph memory, using exposure correction and pre-registered invalidation criteria. Implement ablation experiments that delete or downgrade nodes/edges to measure recovery (deletion-sensitive function) and test LoopGuard-style selective-pruning interventions. When reporting, stratify MDI by error type (e.g. entity resolution failures versus inference errors) to diagnose sources of entrenchment.

中文

采用类似 MDI 的度量来量化图记忆中低置信或无效三元组的累积，使用曝光校正与预注册的失效判定标准。实施删减消融实验以删除或降级节点/边，衡量系统的恢复能力（对删除敏感的功能），并测试 LoopGuard 式的选择性修剪干预。在报告中按错误类型分层 MDI（例如实体解析失败与推理错误），以诊断固化来源。

Future directions

English

Empirically validate HSH in a KG-based agentic RAG system by measuring MDI for stored triples and bridge entities, and test whether immune-like mechanisms reduce entrenchment without harming retrieval accuracy. Extend the MDI concept into a graph-level Homeostatic Drift Integral (HDI) tailored to multi-hop retrieval paths.

中文

在基于知识图谱（KG）的智能体 RAG 系统中对 HSH 进行实证检验：测量存储三元组与桥接实体的 MDI，并测试免疫式机制在降低固化同时是否保持检索精度。将 MDI 概念扩展为适用于多跳检索路径的图级稳态漂移积分（HDI）。

---

06 · LoRA based instruction fine tuning of large language models for agronomic triple extraction and GraphRAG knowledge retrieval

基于LoRA的指令微调大语言模型用于农学三元组抽取与GraphRAG知识检索

Research record	Details
Authors	Rohini Kokare, Sunil B. Mane
Published	2026-10-07
Sources	openalex
Focus	agronomic triple extraction, LoRA instruction fine-tuning, domain-specific LLM, GraphRAG retrieval, semantic retrieval 农学三元组抽取, LoRA 指令微调, 领域特定大模型, GraphRAG 检索, 语义检索

English

Relevant for the thesis as a concrete example of domain-specific LoRA instruction fine-tuning for triple extraction and GraphRAG use, but limited in scope (single domain) and lacking the agentic, persistent graph-memory and entity-resolution investigations central to the thesis.

中文

对论文有用：提供了领域特定 LoRA 指令微调用于三元组抽取并与 GraphRAG 结合的具体实例，但其范围受限（单一领域），且缺乏论文核心的智能体化、持久图记忆与实体消歧研究。

Research synopsis

English

The paper proposes fine-tuning a Llama model with LoRA on an instruction dataset to extract subject–predicate–object triples from unstructured agronomic text (reports, articles, feedback). The fine-tuned Llama-LoRA model is used to generate structured knowledge which is then exposed to a GraphRAG retrieval layer for query answering and semantic search. The authors report improved performance after domain-specific instruction fine-tuning, claiming gains measured by ROUGE and BLEU against other retrieval techniques, and argue that this pipeline mitigates limitations in knowledge representation and unstructured data extraction for the agricultural domain.

中文

本文提出使用 LoRA 在指令数据集上对 Llama 模型进行微调，以从非结构化农学文本（报告、文章、反馈）中抽取主谓宾三元组。微调后的 Llama-LoRA 模型用于生成结构化知识，并通过 GraphRAG 层用于查询检索与语义搜索。作者报告域特定指令微调后性能提升，使用 ROUGE 和 BLEU 与其他检索方法比较并获较好结果，认为该流程缓解了农学领域知识表示与非结构化数据抽取的限制。

Thesis relevance

English

Overlap: both focus on extracting triples and assembling structured knowledge for retrieval and explicitly use GraphRAG-style retrieval. Differences/limitations: the paper is domain-specific (agronomy) and emphasizes LoRA instruction fine-tuning and automatic triple extraction, but it does not describe agentic paradigms, persistent or incrementally accumulated graph memory, entity-resolution strategies, or multi-hop benchmarks like MultiHop-RAG. Complementarity: the fine-tuning and triple-extraction techniques could be incorporated as a component in the thesis pipeline to improve triple quality before graph insertion.

中文

重合点：两者都关注三元组抽取与构建用于检索的结构化知识，并明确使用 GraphRAG 式检索。差异/限制：该论文针对农学领域，侧重 LoRA 指令微调和自动三元组抽取，但未涉及智能体范式、持久或增量构建的图记忆、实体消歧策略，亦未使用如 MultiHop-RAG 之多跳基准。互补性：其微调与三元组抽取方法可作为论文管线的子模块，用以提高插入图中的三元组质量。

Writing and related work

English

Position this paper in related work under domain-specific triple extraction and instruction fine-tuning; note reliance on ROUGE/BLEU for retrieval evaluation, which are limited proxies for multi-hop QA and verifiability. Emphasize differences between single-pass KG construction from fine-tuned extractors and the thesis’s goal of persistent, query-aware graph memory assembled from agentic interactions.

中文

将该论文归入领域特定三元组抽取与指令微调的相关工作；指出其使用 ROUGE/BLEU 评估检索存在局限，这些指标不足以反映多跳问答与可核查性。强调基于微调抽取器的一次性知识构建，与论文旨在通过智能体交互增量构建查询感知图记忆的目标存在差异。

Method and evaluation

English

Consider evaluating the Llama-LoRA extractor as a drop-in triple-extraction module within the thesis system and measure its precision/recall on annotated triples from your MultiHop-RAG documents. Compare downstream effects on multi-hop retrieval and bridge-entity continuity when using this extractor versus your existing extractor. Report extraction confidence and latency, and include ablations that vary fine-tuning data size and instruction prompt formats.

中文

可将 Llama-LoRA 抽取器作为可替换的三元组抽取模块接入论文系统，并在 MultiHop-RAG 的标注三元组上测量其精确率/召回率。比较将该抽取器与现有抽取器用于下游多跳检索和桥接实体连续性时的差异。记录抽取置信度和延迟，并做消融实验，改变微调数据规模与指令提示格式。

Future directions

English

Extend the reported pipeline beyond agriculture and assess triple extractor generalization on multi-domain corpora; evaluate integration with persistent graph memory and explicit entity-resolution strategies; and replace ROUGE/BLEU with retrieval and QA metrics suited for multi-hop, verifiable question answering.

中文

将所述流程扩展到农业之外，评估三元组抽取器在跨域语料上的泛化；研究其与持久图记忆及显式实体消歧策略的整合；并用适合多跳、可核查问答的检索与 QA 指标替代 ROUGE/BLEU。

---