Utilização de Chain-of-Thought Prompting para classificação de questões da OBI

Abstract

The Brazilian Informatics Olympiad (OBI), in its initiation category, evaluates the logical problem-solving abilities of basic education students; however, the study material recommended by the organizers is based on older editions of the competition, which may compromise preparation for more recent exams. This work investigates whether the classes and types defined in the recommended material remain adequate for current editions of the OBI, and analyzes how modern prompting techniques, especially Chain-of-Thought (CoT), can support the automatic classification of these questions. The methodology involved an evaluation of the exams and their questions, followed by the incremental development of different prompts applied to the GPT-4o model, starting from zero-shot and one-shot approaches and progressing to versions with explicit reasoning. The results indicate that the categories proposed in the study material remain suitable for recent exams and highlight the potential of CoT prompting as a support tool for educational analyses and the development of automated question classification systems. The main contribution of this work is the documentation of a prompt-writing methodology that can serve as a reference for future classification research involving Large Language Models (LLM).