跳转至

文章背景与核心概要

本文针对大语言模型（LLM）在电网模型问答中面临的严格上下文预算限制，提出了一种名为“种子锚定图渲染”（seed-anchored graph rendering）的确定性方法。该方法无需调整或训练特定参数，便可优先处理查询局部的图证据，并通过贪婪受限上下文前缀确保保留种子局部的答案渲染单元。

在基于通用电网模型交换标准（CGMES）的通用信息模型（CIM）网络上的评估表明，与朴素方法和基线图检索流水线相比，该方法在证据保留率和问答准确率上均取得了显著提升。在固定 8,000 字符上下文预算下，SmallGrid 拓扑系列的准确率从 \(0.450\) 飙升至 \(0.970\)，同时避免了额外的 LLM 图构建令牌开销。

面向行业标准电网信息与交换模型问答的种子锚定预算受限图渲染 (Seed-Anchored Budget-Bounded Graph Rendering for Question Answering on Industry-Standard Power-Grid Information and Exchange Models)

摘要 (Summary)

本文引入了种子锚定图渲染（seed-anchored graph rendering），这是一种确定性方法，旨在优化大语言模型（LLM）在严格上下文预算下对电网模型的问答。该方法在无需调整或学习参数的情况下，优先处理查询局部的图证据，保证在贪婪受限上下文前缀中保留种子局部的答案承载渲染单元。对使用通用电网模型交换标准（CGMES）的通用信息模型（CIM）网络模型的评估表明，与朴素方法和基线图检索流水线相比，该方法在证据保留和问答准确性方面有显著提升。

This paper introduces seed-anchored graph rendering, a deterministic method designed to optimize large language model (LLM) question answering over power-grid models under strict context budgets. By prioritizing query-local graph evidence without requiring tuned or learned parameters, the method guarantees the preservation of seed-local answer-bearing render units within a greedy bounded-context prefix. Evaluations on Common Information Model (CIM) network models using the Common Grid Model Exchange Standard (CGMES) demonstrate significant improvements in evidence retention and question-answering accuracy compared to naive approaches and baseline graph retrieval pipelines.

元数据 (Metadata)

  • arXiv ID: arXiv:2609.02011 [eess.SY]
  • 作者: Jayakumar Manoharan, Yamini Sehgal
  • 提交时间: 2026年9月2日
  • 主学科: 系统与控制 (eess.SY)
  • 次学科: 人工智能 (cs.AI)、信息检索 (cs.IR)
  • 目标出版物: 投递至 Engineering Applications of Artificial Intelligence
  • DOI: 10.48550/arXiv.2609.02011
  • arXiv ID: arXiv:2609.02011 [eess.SY]
  • Authors: Jayakumar Manoharan, Yamini Sehgal
  • Submitted: September 2, 2026
  • Primary Subject: Systems and Control (eess.SY)
  • Secondary Subjects: Artificial Intelligence (cs.AI), Information Retrieval (cs.IR)
  • Target Publication: Submitted to Engineering Applications of Artificial Intelligence
  • DOI: 10.48550/arXiv.2609.02011

摘要详情 (Abstract)

基于电网模型的大语言模型问答必须遵守固定的上下文预算。我们引入了种子锚定图渲染，这是一种确定性方法，它在不添加超出共享跳数边界和上下文预算的方法特定调整或学习参数的情况下，优先处理查询局部的图证据。

该方法提供了一个可检查的条件，在此条件下，预定义的种子局部答案承载渲染单元被保留在贪婪受限上下文前缀中。我们在通过通用电网模型交换标准（CGMES）交换的通用信息模型（CIM）网络模型上评估了该方法： * 证据保留率 (Evidence Retention): 在两个受预算约束的 CGMES 编码上，朴素的“描述优先”渲染保留了每个单跳项目的局部证据，但多跳项目的保留率仅为 \(0.12\) 和 \(0.00\)，而种子锚定渲染保留了所有此类证据。 * 准确率 (Accuracy): 在预先注册的来自 SmallGrid 拓扑系列的全新 100 项测试库中，在固定 8,000 字符上下文预算下，准确率从 \(0.450\) 提高到 \(0.970\)。 * 对比 (Comparison): 在通用的检索和渲染流水线下，标准原生的种子锚定图匹配或超过了 LightRAG、Microsoft GraphRAG 和 HippoRAG 生成的提取图表示，同时避免了 LLM 图构建令牌的消耗。

(注：结果特定于所评估的 CIM/CGMES 模型、读取器和上下文预算；它们关注的是预算受限检索，而非通用问答。)

Large language model question answering over power-grid models must respect a fixed context budget. We introduce seed-anchored graph rendering, a deterministic method that prioritizes query-local graph evidence without adding method-specific tuned or learned parameters beyond the shared hop bound and context budget.

The method provides a checkable condition under which predefined seed-local answer-bearing render units are preserved in a greedy bounded-context prefix. We evaluate the approach on Common Information Model (CIM) network models exchanged through the Common Grid Model Exchange Standard (CGMES): * Evidence Retention: On two budget-binding CGMES encodings, naive descriptions-first rendering retains local evidence for every single-hop item but only \(0.12\) and \(0.00\) of multi-hop items, whereas seed-anchored rendering retains all such evidence. * Accuracy: On a preregistered fresh 100-item bank from the SmallGrid topology family, accuracy rises from \(0.450\) to \(0.970\) under a fixed 8,000-character context budget. * Comparison: Under a common retrieval and rendering pipeline, the standards-native seed-anchored graph matches or exceeds extracted graph representations produced by LightRAG, Microsoft GraphRAG, and HippoRAG, while avoiding LLM graph-construction tokens.

(Note: The results are specific to the evaluated CIM/CGMES models, reader, and context budget; they concern budget-bounded retrieval rather than general question answering.)