漫步黑暗面:使用 DARKSIDE 进行逻辑增强生成与连贯性审计
文章背景与核心概要
大语言模型(LLM)在模式识别方面表现出色,但在处理连贯论述所需的排除逻辑路径时往往力不从心。当面对复杂的无稽之谈(如虚构权威或隐秘类比)时,标准大语言模型倾向于将这些输入视为基础事实,从而将错误固化到其结构化输出中。
本文介绍了 DARKSIDE,这是一种连贯性审计方法,旨在充当诸如 POLANYI++ 等逻辑增强生成(LAG)系统的控制层。通过将论述的“轨迹”形式化为由累积排除组成的数据结构,并引入用于对指代物进行分类的“担保轴”(Warranted、Unattested、Misattributed 或 Fabricated),DARKSIDE 充当了一道认知防火墙。在 BSBench 对抗性语料库上的实证评估表明,这种架构方法有效地弥合了结构模式与逻辑路径之间的鸿沟,显著提升了大语言模型生成的知识图谱的可靠性。
漫步黑暗面 (Walking on the DARKSIDE)
Authors: Aldo Gangemi, Emanuele Bottazzi
Date: August 24, 2026
Subject: Artificial Intelligence (cs.AI); Logic in Computer Science (cs.LO)
Identifier: arXiv:2608.23370
Authors: Aldo Gangemi, Emanuele Bottazzi
Date: August 24, 2026
Subject: Artificial Intelligence (cs.AI); Logic in Computer Science (cs.LO)
Identifier: arXiv:2608.23370
摘要 (Summary)
大语言模型(LLM)在模式识别方面表现出色,但在追踪连贯论述所需的排除逻辑路径时往往会失效。当遇到复杂的无稽之谈——例如虚构的权威或隐蔽的类比——标准的大语言模型倾向于将这些输入视为有根据的事实,将错误实体化到其结构化输出中。
本文引入了 DARKSIDE,这是一种连贯性审计方法,旨在作为诸如 POLANYI++ 等逻辑增强生成(LAG)系统的导向层。通过将论述的“轨迹”形式化为累积排除的数据结构,并实现一个用于对指代物进行分类的“担保轴”(担保、未证实、错误归类或虚构),DARKSIDE 充当了认知防火墙的角色。在 BSBench 对抗性语料库上的实证评估表明,这种架构方法有效地填补了结构模式与逻辑路径之间的鸿沟,显著提高了 LLM 生成的知识图谱的可靠性。
Large Language Models (LLMs) excel at pattern recognition but often fail to track the logical path of exclusions required for coherent discourse. When presented with sophisticated nonsense—such as fabricated authorities or surreptitious analogies—standard LLMs tend to treat these inputs as grounded facts, reifying errors into their structured outputs.
This paper introduces DARKSIDE, a coherence auditing method designed to act as a steering layer for Logic-Augmented Generation (LAG) systems like POLANYI++. By formalizing the "trail" of discourse as a data structure of accumulated exclusions and implementing a "warrant axis" to classify referents (Warranted, Unattested, Misattributed, or Fabricated), DARKSIDE acts as an epistemic firewall. Empirical evaluation on the BSBench adversarial corpus demonstrates that this architectural approach effectively scaffolds the gap between structural patterns and logical paths, significantly improving the reliability of LLM-generated knowledge graphs.
核心特性 (Key Features)
- 连贯性审计: 实现了一个显式的数据结构,用于在整个论述过程中追踪排除项。
- 担保轴: 一种针对命名指代物的分类系统,当虚构或不受支持的信息超过定义阈值时,会触发将
DelegationRiskAssessment设为UNSAFE(不安全)。 - 认知防火墙: 作为 LLM 前向传递之上的保护层,防止无稽的输入被实体化为扩展知识图谱(XKG)。
- Coherence Auditing: Implements an explicit data structure to track exclusions over the course of a discourse.
- Warrant Axis: A classification system for named referents that triggers a
DelegationRiskAssessmenttoUNSAFEwhen fabricated or unsupported information exceeds defined thresholds.- Epistemic Firewall: Functions as a protective layer over LLM forward passes, preventing the reification of nonsensical inputs into Extended Knowledge Graphs (XKG).
访问与资源 (Access & Resources)
- 全文 PDF: 查看论文
- DOI: 10.48550/arXiv.2608.23370
- 许可协议: Creative Commons BY-NC-SA 4.0
- Full-Text PDF: View Paper
- DOI: 10.48550/arXiv.2608.23370
- License: Creative Commons BY-NC-SA 4.0

元数据 (Metadata)
| 字段 (Field) | 详情 (Details) |
|---|---|
| 评论 (Comments) | 20页,2张图表,数个表格 (20 pages, 2 figures, several tables) |
| 主学科 (Primary Subject) | 人工智能 (Artificial Intelligence, cs.AI) |
| 次学科 (Secondary Subject) | 计算机科学中的逻辑 (Logic in Computer Science, cs.LO) |
| 提交历史 (Submission History) | [v1] 2026年8月24日 星期一 ([v1] Mon, 24 Aug 2026) |
Field Details Comments 20 pages, 2 figures, several tables Primary Subject Artificial Intelligence (cs.AI) Secondary Subject Logic in Computer Science (cs.LO) Submission History [v1] Mon, 24 Aug 2026