‘A last-minute challenge became my biggest breakthrough’
Data Science & AI student Nataliia Bagan combines a passion for mathematics, language, and artificial intelligence. Her exceptional bachelor’s thesis on improving reasoning in large language models earned her a nomination for the Leiden Science Young Talent Award 2025.
How did you come to choose this bachelor?
‘I fell in love with mathematics in high school. I was drawn to its logic and technical rigour, but I also wanted to apply those skills in a more practical way. This led me to pursue Informatics, and I even qualified for the All-Ukrainian Girls’ Olympiad—although it never took place because of the invasion of Ukraine. The war forced me to flee to the Netherlands, where I discovered the Data Science & AI programme in Leiden. The combination of human cognition and computational systems fascinated me immediately. This was still before AI gained global popularity through tools like ChatGPT. The field has kept me deeply engaged ever since.’
How did you come up with the topic for your thesis?
During my Bachelor’s, I became passionate about Natural Language Processing (NLP): teaching computers to understand human language. Linguistics have always been a personal interest of mine, so I was thrilled to find a way to integrate it with my technical studies. I joined the group of professor Zhaochun Ren, who works on a broad range NLP topics. The specific topic for my thesis allowed me to combine NLP with my other favourite subject: Reinforcement Learning, in this method, models improve their performance through trial and error and feedback.’
What did you research and what did you discover?
‘I worked on reasoning in Large Language Models. For complex questions, these models often need to break them into smaller sub questions. Computer scientists are always looking for ways to make that process more accurate and efficient. I showed that Reinforcement Learning, specifically Monte Carlo Tree Search (MCTS, see box), can lead to more reliable answers. I developed a simplified experimental approach that could achieve the same strong performance as existing, much more complex MCTS methods.’
How does it work? – Monte Carlo Tree Search (MCTS)
Monte Carlo Tree Search is a method that helps a computer make smart decisions by exploring possible actions and their outcomes. It simulates many random possibilities, sees which actions usually lead to the best results, and then chooses the most promising one.
What was a memorable moment while working on your thesis?
‘Overnight, my method became outdated’
‘Two weeks before my deadline, the benchmark model I was comparing against released an updated version—one that suddenly outperformed my approach. Overnight, my method seemed outdated. Instead of submitting weaker results, I pushed harder and spent two intense weeks running new experiments. Three days before the deadline, I finally found an even better-performing method. Stressful, but ultimately incredibly rewarding.’
What are your plans for the future?
‘My ultimate goal is to become a university professor. I believe we need more researchers who are not only strong academically but also passionate about teaching and genuinely connecting with students. To be an inspiration just by genuinely enjoying teaching. My next steps are to pursue a Master’s and then a PhD to build the expertise needed for this path.’