Research Group
Kristína Okasová
Research areas: natural language processing, large language models, reinforcement learning, memory mechanisms of LLMs
Position: PhD Student
Kristína is a PhD student in the NLP team, focusing on limitations of LLMs, reinforcement learning, and memory mechanisms in LLMs. She holds both a Bachelor’s degree in Informatics and a Master’s degree in Intelligent Software Systems from Slovak University of Technology. Her interest in artificial intelligence began early in her studies, initially through machine learning research for her Bachelor’s thesis — work that later led to a published article and first sparked her interest in pursuing a research career. Over time, her curiosity developed into a strong focus on deep learning and natural language processing, and for her Master’s thesis, she explored fine-tuning LLMs.
Currently, she is completing the first year of her PhD programme with the dissertation topic Addressing Limitations of LLMs. Her main research interests are reinforcement learning, long-context handling in LLMs, and their enhancement through memory mechanisms. In the coming years, she aims to further develop these areas, particularly focusing on their application to small language models and low-resource languages. Moreover, her ambition is to attend NLP-oriented summer schools, NLP-focused conferences, and an internship abroad in order to present her research and connect with other researchers working in the same field.
PhD topic: Addressing Limitations of Large Language Models
Supervising team: Michal Gregor (KInIT)
Large language models (LLMs) are a powerful tool driving recent progress in artificial intelligence (AI). They integrate broad world knowledge into AI systems, follow natural language instructions, perform tasks in few-shot settings thanks to in-context learning, and increasingly support multiple modalities such as images and audio.
Despite these advances, LLMs face limitations that hinder their safe and widespread use. They tend to produce unsupported outputs (hallucinations), struggle with multi-step reasoning and planning, and show weaknesses in multimodal integration, such as recognizing fine-grained visual concepts. Moreover, they acquire new knowledge and skills with far lower sample efficiency than humans, which poses particular challenges for low-resource languages.
