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

The researchers proposed a framework called DETECT-REMASK-REPAIR, aimed at addressing the issue of outdated summary information in dynamic contexts. This framework utilizes a mask diffusion language model to locally repair outdated parts of existing summaries while retaining supportive content. Experiments show that this method effectively improves the quality of early drafts on DialogSum and StreamSum (synthetic event timeline benchmarks). The cost of single-step repairs is less than 0.5 seconds, achieving a balance between fidelity, speed, and preservation. Additionally, this framework can be used as a post-processing step to enhance the fidelity of autoregressive systems.