Institutional Repository of Vilnius University
Didžiųjų kalbos modelių analizė algoritmų problemų sprendimui programų inžinerijoje /
Abstract
dc:descriptionThis study presents an automated testing and analysis system for five large language models (GPT‐4o, GPT‐4.1 Mini, Gemini 1.5 Pro, Gemini 2.5 Pro, and DeepSeek Reasoner), which automatically generated, stored, and executed 2520 responses to 168 algorithmic programming tasks within a Pyt‐ hon environment. The research focused on how three prompting strategies—zero‐shot, few‐shot, and chain‐of‐thought—affect model accuracy, response time, and error rate across different levels of task complexity. The comparative analysis revealed that Gemini 2.5 Pro maintained the highest and most stable accuracy, while Gemini 2.5 Pro and DeepSeek Reasoner improved most significantly when using few‐shot prompts. Conversely, GPT‐4o demonstrated the greatest accuracy degrada‐ tion on the most difficult tasks, which was partially mitigated by the chain‐of‐thought method at the medium difficulty level.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zubavičius, Justas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD239032822&prefLang=en_US