Supervising Team: Michal Gregor (KInIT)

Kristína’s research explores certain core limitations of large language models (LLMs) – such as weaknesses in reasoning abilities, execution, long-context capabilities and memory, or data-efficient learning. Even though LLMs now possess broad capabilities and acquire extensive general knowledge during pretraining, they still struggle in these and other areas. Kristína’s work focuses on understanding these limitations and on developing training strategies, architectures, evaluation methods, and other mechanisms to mitigate them, hoping to ultimately help make LLMs – including their smaller, more efficient variants – more capable, reliable, and data-efficient.