An LLM-driven method for open-domain Timeline Summarization: rather than being handed a curated news corpus, the model builds a dated chronological summary of a topic through iterative self-questioning. Each round it reflects on how the events gathered so far are causally linked, poses new questions, retrieves documents that answer them (from the open web or an offline knowledge base), and refreshes both the event set and the timeline. The work also introduces Open-TLS, a dataset of recent-news timelines authored by professional journalists to evaluate open-domain TLS under information overload, and shows the approach rivals state-of-the-art closed-domain systems that receive a prepared corpus.
Timeline Summarization (TLS) is the task of turning a large, scattered set of event-related documents about a news topic into a coherent, dated chronological summary - effectively building an event g…