Detect, Remask, Repair: Diffusion Editing for Faithful Summarization of Evolving Contexts

The article introduces the DETECT-REMASK-REPAIR framework, which employs masked diffusion language models to update outdated spans in existing summaries while preserving supported content. It presents the StreamSum benchmark for evaluating evolving-context summarization and demonstrates that localized diffusion repair offers a more efficient alternative to full summary regeneration, achieving significant reductions in repair time and enhancing the faithfulness of summaries. This framework is particularly relevant for practitioners looking to improve the accuracy of summaries in dynamic contexts without the overhead of complete rewrites.