Multi-agent for Dynamic-Granularity Timeline Summarization

Dynamic-granularity Timeline Event Summarization (DTELS), a specialized variant of Timeline Summarization, is pivotal for distilling coherent narratives from large text corpora under flexible granularity constraints. Despite its importance, existing methods struggle to balance narrative coherence with adaptive granularity. To address this gap, we introduce a novel multi-agent framework that revolutionizes DTELS by dividing the task into three distinct stages: article grouping, event extraction, and summary generation. Our approach leverages predefined rules and a collaborative multi-agent architecture to enhance information extraction accuracy and ensure granular consistency across diverse datasets. Empirical evaluations on the newly curated DTELS-Bench demonstrate that our framework significantly outperforms state-of-the-art methods, achieving superior narrative coherence and granularity adaptability. Furthermore, our method achieved second place in the CCKS 2025 Shared Task on Event Timeline Generation for Social Media, validating its effectiveness in real-world scenarios. Our method nearly achieved perfect scores in granularity and factuality, while achieved 0.2 informativeness score, a 20% enhancement over other top-ranking methods. These advancements establish a new benchmark for timeline summarization, offering robust tools for applications requiring dynamic event summarization.