Attribution by Construction: Claim-Anchored Evidence for Faithfulness-Oriented Multi-Document Summarization

Multi-document summarization (MDS) often suffers from two primary issues: hallucinations in generated text and coarse, post-hoc citations that make it difficult for readers to verify specific claims. The author investigates whether a modular, claim-anchored pipeline can improve factual faithfulness and traceability compared to end-to-end large language models (LLMs) that typically flatten source information and obscure multi-source support.

The author introduces CAMS (Claim-Anchored Multi-document Summarization), a framework that replaces end-to-end generation with a five-stage pipeline:

1. Claim Extraction: Decomposes source documents into atomic, decontextualized claims while maintaining token-level provenance (source spans).
2. Clustering & Conflict Detection: Groups equivalent claims across documents and identifies contradictions using NLI models.
3. Content Selection: Uses a gradient-boosted decision tree to select a salient, support-aware subset of claims, allowing for a controllable trade-off between coverage and faithfulness.
4. Attributable Rewriting: Generates summary sentences that are explicitly anchored to the selected claims.
5. Verification: Performs a two-way check (support and no-overflow) to ensure generated sentences are backed by the cited source spans, with optional repair retries.

CAMS matches the summary quality of strong end-to-end baselines while significantly improving faithfulness and citation precision. By explicitly managing cross-document redundancy and conflict, the framework increases multi-source attribution accuracy by approximately two-thirds. The modular design also exposes a controllable trade-off between summary coverage and factual faithfulness, which is typically implicit in end-to-end models.

This work demonstrates that structured intermediate representations—specifically atomic claims linked to source spans—can bridge the gap between fluent abstractive summarization and rigorous, verifiable attribution. By decoupling content selection from realization, CAMS provides a more interpretable and controllable path for building reliable summarization systems in multi-source settings.