| Abstract [eng] |
The master's thesis addresses the challenge of comprehending mathematical word problems by applying large language models (LLMs). The goal of the thesis is to create an AI system that would not simply provide answers but would help students understand the core of the problem. Using the research methodology, a hybrid system was designed and implemented, integrating the Google Gemini model with calculation tools (MathJS) and a system prompt. This solution successfully eliminated the hallucinations characteristic of generative models and ensured a pedagogically sound Socratic dialogue in Lithuanian. An experimental evaluation with 48 respondents confirmed the high technological acceptability of the system, achieving a SUS score of 79.43. As many as 95.83% of participants subjectively indicated that interactive dialogue helped them to independently understand the logic of solving the task, rather than just obtaining the result. The results of the study prove that specialized AI agents can act as effective learning partners, promoting deep rather than superficial learning. |
