Dynamic agenda-aware real-time meeting summarization with large language models

Abstract Meeting summarization has become increasingly important with the rise of remote collaboration. However, existing approaches mostly generate summaries post hoc and struggle with the real-time and dynamic nature of multi-topic meetings. In this work, we define the task of dynamic agenda-aware real-time meeting summarization, which requires incrementally producing coherent summaries as a meeting progresses, guided by an evolving agenda. We identify key challenges: (1) long, multi-participant dialogues covering changing topics, (2) the need to update summaries with new information while minimizing redundancy, and (3) aligning summaries with agenda items that may shift during the meeting. To address these issues, we propose a novel summarization framework that integrates fine-tuned large language models (LLMs) with an explicit agenda tracking mechanism. Our method leverages a pre-trained LLM and incorporates the meeting agenda as contextual guidance, dynamically adjusting the focus of the summary as the discussion moves through agenda items. Experimental results on three public meeting datasets demonstrate that our approach achieves higher ROUGE-1, ROUGE-2, ROUGE-L, BERTScore, and METEOR scores than competitive methods. The proposed system generates high-quality real-time summaries, offering participants timely insights and agenda-aligned highlights throughout the meeting. These results signify a step towards more interactive and intelligent meeting assistants, and the techniques presented can be generalized to other streaming summarization scenarios. The contributions of this work open new directions for online and interactive summarization research, bridging the gap between offline summary generation and the practical need for live meeting minutes.