跳转至

文章背景与核心概要

端到端大语言模型（LLM）在生成多文档摘要时表现出流利性，但往往面临严重的幻觉问题，且通常只提供粗粒度的后验归因（如整篇文档或段落），导致单个陈述难以验证。本文重新审视了模块化的“提取-选择-重写”（Extract–Select–Rewrite）范式，并引入了 CAMS（Claim-Anchored Multi-document Summarization，声明锚定多文档摘要） 框架。

CAMS 在内容生成（实现）之前进行内容定位：首先提取带有 Token 级溯源的原子声明，聚类等效声明并标记冲突，选择支持感知（support-aware）的子集，最后将其重写为可验证的摘要。在 MultiNews、DiverseSumm 和 WCEP 等数据集上的评估表明，CAMS 在保持摘要质量的同时，大幅提升了忠实度与引用精度，将多源归因准确率提高了约三分之二。

Faithful by Construction: Claim-Anchored Attribution for Multi-Document Summarization

Summary

End-to-end Large Language Models (LLMs) often suffer from hallucinations and offer only coarse, post hoc attributions (such as whole documents or passages), making individual statements difficult to verify. This paper revisits the modular Extract–Select–Rewrite paradigm and introduces CAMS (Claim-Anchored Multi-document Summarization). CAMS localizes content before realization by extracting atomic claims with token-level provenance, clustering equivalent claims while flagging conflicts, selecting a support-aware subset, and rewriting the output into a verifiable summary. Evaluated across datasets like MultiNews, DiverseSumm, and WCEP, CAMS substantially improves faithfulness and citation precision while maintaining summary quality.

端到端大语言模型（LLM）通常饱受幻觉问题困扰，且只能提供粗粒度的后验归因（如整篇文档或段落），导致单个陈述难以验证。本文重新审视了模块化的提取-选择-重写（Extract–Select–Rewrite）范式，并引入了 CAMS（声明锚定多文档摘要，Claim-Anchored Multi-document Summarization）。CAMS 在内容实现之前进行内容定位：通过提取具有 Token 级溯源的原子声明、聚类等效声明并标记冲突、选择支持感知（support-aware）的子集，最后将输出重写为可验证的摘要。在 MultiNews、DiverseSumm 和 WCEP 等数据集上的评估表明，CAMS 在保持摘要质量的同时，大幅提升了忠实度与引用精度。

Metadata

  • arXiv ID: arXiv:2606.23989
  • Primary Subject: Computation and Language (cs.CL)
  • Secondary Subjects: Artificial Intelligence (cs.AI)
  • Author: Shuo Guan
  • Submitted on: 22 June 2026
  • Last Revised: 3 September 2026 (Version v3)
  • DOI: 10.48550/arXiv.2606.23989
  • arXiv ID: arXiv:2606.23989
  • Primary Subject: Computation and Language (cs.CL)
  • Secondary Subjects: Artificial Intelligence (cs.AI)
  • Author: Shuo Guan
  • Submitted on: 22 June 2026
  • Last Revised: 3 September 2026 (Version v3)
  • DOI: 10.48550/arXiv.2606.23989

Abstract

End-to-end large language models (LLMs) produce fluent multi-document summaries but remain prone to hallucination, and the attributions they offer are typically coarse (whole documents or passages) and generated post hoc, leaving each summary statement hard to verify.

端到端大语言模型（LLM）能够生成流利的多文档摘要，但依然容易产生幻觉，且它们提供的归因通常很粗糙（针对整篇文档或段落）且属于事后生成，这使得摘要中的每个陈述都难以验证。

We revisit the modular Extract–Select–Rewrite paradigm and recast its intermediate representation as the unit of attribution. We present CAMS, a Claim-Anchored Multi-document Summarization framework that: 1. Extracts atomic claims with token-level provenance from every source document. 2. Clusters equivalent claims across documents while flagging inter-source conflicts. 3. Selects a support-aware and salient subset. 4. Rewrites the selection into a summary in which every sentence is anchored to a support-checked claim that links back to one or more source spans.

我们重新审视了模块化的“提取-选择-重写”范式，并将其中间表示重塑为归因的单位。我们提出了 CAMS（声明锚定多文档摘要框架），该框架能够： 1. 提取：从每个源文档中提取带有 Token 级溯源的原子声明。 2. 聚类：对跨文档的等效声明进行聚类，同时标记不同来源之间的冲突。 3. 选择：选择一个支持感知且显著的子集。 4. 重写：将所选内容重写为摘要，其中每个句子都锚定到一个经过支持检查的声明上，该声明可追溯到一个或多个源文本片段。

Because content is localized before it is realized, the pipeline is attribution-oriented by construction and faithfulness-oriented by construction: it structurally preserves fine-grained, multi-source traceability while using support-aware selection, constrained rewriting, and verification to encourage, rather than guarantee, factual faithfulness.

由于内容在实现之前就已经完成了定位，因此该流水线在结构上就是面向归因和面向忠实度的：它在结构上保留了细粒度的多源可追溯性，同时利用支持感知选择、受限重写和验证来促成（而非绝对保证）事实忠实度。

We evaluate quality, faithfulness, and localization on MultiNews, analyze conflict handling on DiverseSumm, and test zero-shot transfer on WCEP, using a two-regime protocol that separates reference-free citation quality from gold-aligned localization accuracy. Additionally, we add an evaluator-decoupled audit that tests citation precision with a support model never used for selection or verification. CAMS matches strong end-to-end and span-attribution baselines on summary quality while substantially improving faithfulness and citation precision, lifting multi-source attribution accuracy by roughly two-thirds, and exposing a controllable faithfulness–coverage trade-off that end-to-end models leave implicit.

我们在 MultiNews 上评估了质量、忠实度和定位能力，在 DiverseSumm 上分析了冲突处理，并在 WCEP 上测试了零样本迁移。我们采用了一种两阶段协议，将无参考引用质量与黄金标准对齐的定位准确率区分开来。此外，我们引入了一项与评估器解耦的审计，使用一个从未用于选择或验证的支持模型来测试引用精度。CAMS 在摘要质量上与强大的端到端和片段归因基线相媲美，同时显著提高了忠实度和引用精度，将多源归因准确率提升了约三分之二，并揭示了端到端模型所含糊不清的可控“忠实度-覆盖率”权衡。

Access the Paper & Resources

External Tools & Integrations

外部工具与集成