NASK Researcher Awarded SONATA 21 Grant for AI Safety Research Project
Among this year’s successful applicants is a researcher from NASK. Funding of PLN 1,174,946 has been awarded to the project “Eval FM: Reliable Evaluation of Safety in AI Models”
The project is led by Sebastian Cygert, Head of the AI Safety and Transparency Team at NASK.
The SONATA programme supports researchers at the early stages of their independent scientific careers. In this year’s edition, NCN selected the most outstanding research proposals from across Poland for funding.
The Eval FM project addresses one of the most important challenges in modern artificial intelligence: how to evaluate and develop AI models so that they are not only effective, but also safe, trustworthy, and robust.
The research will focus on three key areas:
- assessing whether AI models genuinely generalize to new situations rather than simply memorizing training data,
- identifying hidden vulnerabilities and mechanisms that may lead to unpredictable or unsafe system behavior,
- developing fine-tuning methods that preserve safety while improving model performance and usefulness.
As part of the project, the research team will develop novel diagnostic tools and evaluation methods that enable more reliable assessment of AI models. The team will also work on solutions for detecting so-called “backdoors” and reducing the risk of safety degradation when adapting models to specific applications.
The project’s outcomes may support the safe deployment of artificial intelligence in areas of significant social importance, including healthcare, public administration, education, media analysis, and cybersecurity, where reliability and predictability are essential.
We warmly congratulate Sebastian Cygert and the entire research team. This funding is a testament to the high quality of AI research conducted at NASK and an important contribution to the development of safe, trustworthy, and responsible artificial intelligence technologies.