Defesa de mestrado do discente Fernando Zimerman, dia 12/08, às 15:00.

Defesa de mestrado do discente Fernando Zimerman, dia 12/08, às 15:00.

Título: Escalando a Criação de Material Didático para Programação Introdutória com Geração Automática de Questões via LLMs

Abstract: Teaching programming, particularly the Introduction to Programming course, poses numerous pedagogical challenges. One of the main challenges is the development of high-quality instructional materials, such as practical programming exercises with sufficient quantity and diversity, especially considering the significant increase in student enrollment in these courses. However, manually creating such exercises is a demanding task that is difficult to scale. Given that current technologies have become essential resources in both academic and professional contexts, this work investigates an alternative approach to address this challenge.
This dissertation proposes the adaptation of the 8-billion-parameter open-source Large Language Model (LLM) Llama~3.1 for the automatic generation of programming exercises, structured according to the syllabus of the Computer Programming I (BCC701) course at the Federal University of Ouro Preto, which serves as the case study for this research. The proposed methodology combines prompt engineering and fine-tuning using QLoRA for parameter adaptation, supported by a dataset comprising 126 practical programming exercises applied in the course between 2021 and 2024.
Furthermore, this work presents an automated approach for evaluating the generated exercises by adopting the LLM-as-a-Judge paradigm with the Gemini~2.5 Flash model. The experimental results demonstrate improvements in both the structural compliance of the generated exercises and their alignment with the requested topic, reaching 100% and 67.56%, respectively, using the proposed strategy. Regarding the criteria assessed through the LLM-as-a-Judge approach, the fine-tuned model achieved improvements of 10% in pedagogical effectiveness and 4% in relevance compared with a zero-shot approach. The results indicate that the proposed methodology is effective in specializing large language models for domain-specific educational contexts, reducing instructors' workload while improving the quality of automatically generated instructional materials.

Banca Avaliadora: Prof. Dr. Pedro Henrique Lopes Silva - UFOP; Prof. Dr. Reinaldo Silva Fortes - UFOP; Prof. Dr. Guilherme Tavares de Assis - UFOP; Prof. Dr.  Leandro Silva Galvão de Carvalho - UFAM

Data: 12/08/2026 Horário: 15:00

Link: https://meet.google.com/oon-odow-bah