Detalhes do Documento

Collaborative problem-solving with LLM: a multi-agent system approach to solve complex tasks using autogen

Autor(es): Barbosa, Ricardo ; Santos, Ricardo ; Novais, Paulo

Data: 2025

Identificador Persistente: https://hdl.handle.net/1822/95211

Origem: RepositóriUM - Universidade do Minho

Projeto/bolsa: info:eu-repo/grantAgreement/FCT/6817 - DCRRNI ID/UIDB%2F04728%2F2020/PT; info:eu-repo/grantAgreement/FCT/6817 - DCRRNI ID/UIDP%2F04728%2F2020/PT; info:eu-repo/grantAgreement/FCT/6817 - DCRRNI ID/UIDB%2F00319%2F2020/PT;

Assunto(s): Autogen; Collaboration; Complex Tasks; Large Language Models; Multi-agent Systems


Descrição

This paper explores the utilization of Large Language Models (LLMs) in Multi-Agent Systems (MAS) in scenarios where the agents are expected to collaborate and negotiate their preferences, creating temporary alliances to achieve a common goal (complex task). MAS have been acknowledged for their potential in facilitating collaboration to solve complex problems. However, widespread adoption of MAS is impeded by challenges related to defining communication languages and developing frameworks that balance specificity for complex use cases with general applicability across different domains. The emergence of LLMs, such as GPT-4, presents a novel approach to MAS, offering advanced natural language processing capabilities that (potentially) circumventing the need for explicit communication language definitions. This paper proposes an MAS implementation utilizing LLMs within the Autogen framework, emphasizing collaboration and negotiation among agents, through a case study involving a product manufacturing scenario where agents are tasked with intricate decision-making. Results from three test scenarios demonstrate the efficacy of this approach, that can be used to enhance further developments in MAS scenarios of application. However, despite the promise, challenges remain, including the cost of running LLMs and the need for further exploration of their capabilities.

This work has been supported by national funds through FCT – Fundação para a Ciência e Tecnologia (Portuguese Foundation for Science and Technology) through the Projects UIDB/04728/2020, UIDP/04728/2020, and the Ricardo Barbosa doctoral Grant with the reference UI/BD/154187/2022. The work of Paulo Novais is supported through the Project UIDB/00319/2020.

Tipo de Documento Comunicação em conferência
Idioma Inglês
Contribuidor(es) Universidade do Minho
facebook logo  linkedin logo  twitter logo 
mendeley logo

Documentos Relacionados