A Multi-Agent LLM-based Approach to Personalized Generation of Learning Activities

• Grant number: 25/10403-2
• Support Opportunities: Regular Research Grants
• Start date: March 01, 2026
• End date: February 28, 2029
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Julio Cesar dos Reis
• Grantee: Julio Cesar dos Reis
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• City of the host institution: Campinas
• Associated researchers: Ahsan Morshed ; André Gomes Regino ; Eryck Pedro da Silva ; Rodolfo Jardim de Azevedo ; Rodrigo Bonacin

Abstract

The creation of innovative technological solutions that enhance teaching and lear-ning is a crucial challenge with the potential to positively impact society. The development ofquestionnaires plays a significant role in fostering students' continuous learning by encouraginginteraction with study content and enabling the assessment of their knowledge level. This pro-ject addresses the challenge of automatically generating multiple-choice questions from inputtexts using Generative Artificial Intelligence. This challenge is heightened by the need to workwith Portuguese texts, which are still underrepresented in current language models. The propo-sed system will allow users, whether teachers or students, to customize the process by selectingthe base text, the number of questions, their types, and difficulty levels. We propose a solu-tion based on a multi-agent system using LLMs for personalized generation of multiple-choicequestions. The project aims to study and propose methods for capturing and representing dataextracted from student activities, which are essential for questionnaire personalization. It willinvolve designing and implementing LLM agents specialized in question generation, conside-ring various question types. We will develop and evaluate techniques to define the difficultylevels of questions, addressing an open problem in the literature with precise calibration foreasy, medium, and hard levels. Additionally, methods will be studied to assess the quality andcorrectness of artificially generated questions. The expected outcomes include innovative so-lutions for generating personalized questionnaires that positively impact learning experiencesand assessment activities in educational environments. We aim to contribute new computationalsolutions that assist instructors in efficiently and cost-effectively creating such activities. Theanticipated results have the potential to advance educational digital technologies, making theteaching process more dynamic and accessible for both students and teachers. (AU)