Scholarlink

<p>The prevalence of algorithm-driven personalization often traps users in information cocoons, creating an urgent need for tools that can break these constraints by synthesizing diverse perspectives. Inspired by the cognitive principles of distributed intelligence and social collaboration, this article presents a novel framework for multi-stance opinion summarization. Specifically, we first propose a novel large language model (LLM) agent collaboration framework for summarizing public opinions from multiple stances in the form of multi-turn conversations, where the LLM as controller and different agents as sub-tasks executors. Furthermore, fusing the online text content and visual multi-modal information based on image description to prompt LLM generate a comprehensive and stance-aware summary about public opinions. Finally, the efficacy of the overall framework is demonstrated through quantitative experiments on stance classification and extensive qualitative case studies on trending topics, confirming its strong capability in understanding and summarizing complex online public opinions. The results showcase a cognitively-inspired approach to information condensation, providing technical support for balanced decision-making for both individuals and organizations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Simulating Collaborative Cognition: an LLM Agent Framework for Multi-Stance Public Opinion Summarization

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Guoshuai Zhang,
  • Jiaji Wu,
  • Gwanggil Jeon,
  • Penghui Wang,
  • David Camacho

摘要

The prevalence of algorithm-driven personalization often traps users in information cocoons, creating an urgent need for tools that can break these constraints by synthesizing diverse perspectives. Inspired by the cognitive principles of distributed intelligence and social collaboration, this article presents a novel framework for multi-stance opinion summarization. Specifically, we first propose a novel large language model (LLM) agent collaboration framework for summarizing public opinions from multiple stances in the form of multi-turn conversations, where the LLM as controller and different agents as sub-tasks executors. Furthermore, fusing the online text content and visual multi-modal information based on image description to prompt LLM generate a comprehensive and stance-aware summary about public opinions. Finally, the efficacy of the overall framework is demonstrated through quantitative experiments on stance classification and extensive qualitative case studies on trending topics, confirming its strong capability in understanding and summarizing complex online public opinions. The results showcase a cognitively-inspired approach to information condensation, providing technical support for balanced decision-making for both individuals and organizations.