PhD in Artificial Intelligence at KInIT: Focused on excellence

The recommender systems are an integral part of almost every modern Web application. Personalized, or at least adaptive, services have become a standard that is expected by the users in almost every domain (e.g., news, chatbots, social media, or search).

Obviously, personalization has a great impact on the everyday life of hundreds of million users across many domains and applications. This results in a major challenge – to propose methods that are not only accurate but also trustworthy and fair. Such a goal offers plenty of research opportunities in many directions:

• Trustworthy recommendation methods for sequence and session-based applications
• Multi-objective and multi-stakeholder recommender systems
• Explaining recommendations
• Fairness and justice in recommendations
• Biases in the recommendations

There are several application domains where these research problems can be addressed, e.g., search, e-commerce, social networks, news, and many others.

Relevant publications:

• V. Bogina, T. Kuflik, D. Jannach, M. Bielikova, M. Kompan, C. Trattner. Considering temporal aspects in recommender systems: a survey. User Modeling and User-Adapted Interaction, 1-39, 2022. https://doi.org/10.1007/s11257-022-09335-w 
• K. Marcinc̆inová, A. Gavornik, M. Mesarc̆ík, M. Kompan Towards Trustworthy Multi-stakeholder Recommender Systems. European Workshop on Algorithmic Fairness, 446-451, 2025. https://proceedings.mlr.press/v294/marcinc-inova25a.html 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Michal Kompan

Lead researcher, KInIT

Peter Brusilovsky

Professor, University of Pittsburgh, USA

Branislav Kveton

Principal Scientist, Adobe Research, USA

Peter Dolog

Associate Professor, Aalborg University, Denmark

Despite the impressive performance of large language models (LLMs), these models are still not well understood. A lot of effort is dedicated to evaluating the capabilities of LLMs and creating guardrails to prevent potentially harmful behaviours. However, small prompt variations still significantly alter the behaviour of LLMs, they often fail in tasks that are similar to the ones seen in training data but contain a small change, or they can be easily used to generate confident misinformation. As the large language models increasingly mediate access to information and are increasingly used by expert and non-expert users, it is essential to develop rigorous methods for evaluating their true capabilities and identifying the sources of their vulnerabilities.

The goal of this research is to build a deeper and more reliable understanding of the LLM behaviour. The first possible research direction involves designing a benchmark that can accurately measure both advanced capabilities (such as reasoning, planning, understanding, math or multilingual capabilities) and common failure modes and vulnerabilities (including hallucination, brittleness, tendency or ease of generating problematic content such as misinformation). An important focus should be on making the benchmark robust to the current issues of data contamination and minimising the resources needed to run it. The second possible research direction leverages mechanistic interpretability to study the internal structures that drive the LLM behaviours. These tools will be used to investigate why specific vulnerabilities arise, how the capabilities arise, and whether targeted interventions can reduce failures while maintaining core capabilities.

By integrating robust behavioural benchmarking with mechanistic insights, the project aims to produce a more principled understanding of how LLMs work and how they can be made safer and more reliable. The application domain includes (but are not limited to) low-resource tasks,  multilingual understanding, misinformation and other problematic behaviour detection (narrative detection) or even LLM user simulation.

Relevant publications:

• Vykopal, I., Pikuliak, M., Srba, I., Moro, R., Macko, D., and Bielikova, M., 2023. 2024. Disinformation Capabilities of Large Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14830–14847, Bangkok, Thailand. Association for Computational Linguistics. https://aclanthology.org/2024.acl-long.793/ 
• Zugecova, A., Macko, D., Srba, I., Moro, R., Kopal, J., Marcinčinová, K., and Mesarčík, M., 2025. Evaluation of LLM Vulnerabilities to Being Misused for Personalized Disinformation Generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 780–797, Vienna, Austria. Association for Computational Linguistics. https://aclanthology.org/2025.acl-long.38/

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Mária Bieliková

Lead researcher, KInIT

Branislav Pecher

Addressing small labeled data in machine learning training

Despite the proliferation of large language models (LLMs) with their proclaimed universal applicability, many tasks of automatic text processing remain insufficiently solved. Tasks that need to be solved in low-resource languages are solved with lower success. Tasks that require sensitivity to details, context, or fresh information also remain problematic. Examples at the intersection of both these dimensions include processing data from social media for the purpose of detecting narratives, analyzing discourse, detecting toxic content, detecting organized influence campaigns, revealing manipulation, supporting fact-checking, or auditing social media recommenders.

Since the above problem can be understood as a lack of data for specific tasks, one of the solutions can be to use the capabilities of LLMs, but not directly as task solvers, but rather for generation or augmentation of data samples. Using these, specialized (smaller) models could be created for specific tasks. The advantage of such an approach is a higher degree of control over the final model (we can control the generated data), as well as a lower price and footprint of the final model (it is smaller). On the other hand, the disadvantage of such an approach is its lower straightforwardness and methodological ambiguity.That ambiguity, however, offers an opportunity for research.

The topic of the dissertation envisions research of methods and methodologies for creating automatic approaches for classification (or other automatic processing) of texts in heterogeneous and unstable domains. These are characterized by a lack of resources,  primarily labeled data, but secondarily also limited computing power.

Relevant publications:

• Cegin, J., Simko, J. and Brusilovsky, P., 2023. ChatGPT to Replace Crowdsourcing of Paraphrases for Intent Classification: Higher Diversity and Comparable Model Robustness. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing https://aclanthology.org/2023.emnlp-main.117/ 
• Cegin, J., Pecher, B., Simko, J., Srba, I., Bieliková, M. and Brusilovsky, P., 2024, August. Effects of diversity incentives on sample diversity and downstream model performance in LLM-based text augmentation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics https://aclanthology.org/2024.acl-long.710/ 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Jakub Šimko

Lead researcher, KInIT

Ján Čegiň

Machine learning with human in the loop

The multilingual capabilities of large language models (LLMs) enable us to generate texts in many languages, or even to analyze the existing multilingual texts. However, there are missing standard methods and metrics to measure the quality of multilingual texts. Analysis, evaluation, and annotation of such texts by humans is costly and time-consuming, while the replicability of the annotations is questionable (due to ensuring the annotators consistency). In some languages, it is quite difficult to obtain human annotations, while also balancing the level of expertise and demographic diversity of annotators across languages. For a while, the researchers are experimenting with using LLMs themselves to evaluate various aspects of text quality (e.g., coherence, grammar correctness, linguistic acceptability) in multiple languages, whether the texts are generated by language models or written by humans. Even the LLM annotations can be biased due to internal biases of the meta-evaluation models caused by training data.

It is important to continually increase the robustness of meta-evaluation (to limit biases and increase objectivity), whether by increasing the number of evaluation LLMs or by ensuring diversity via LLM (person-like) behavior modifications. The feasibility of the meta-evaluation method must be considered by assessing the computational efficiency and overall practical usability. Also, it is required to measure the correlation between meta-evaluation and human judgement and to evaluate for which aspects of text quality and in which languages is such meta-evaluation a suitable and usable alternative.

Relevant publications:

• Macko, D., Kopal, J., Moro, R., and Srba, I., 2025. MultiSocial: Multilingual Benchmark of Machine-Generated Text Detection of Social-Media Texts. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 727–752, Vienna, Austria. Association for Computational Linguistics. https://aclanthology.org/2025.acl-long.36/
• Zugecova, A., Macko, D., Srba, I., Moro, R., Kopal, J., Marcinčinová, K., and Mesarčík, M., 2025. Evaluation of LLM Vulnerabilities to Being Misused for Personalized Disinformation Generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 780–797, Vienna, Austria. Association for Computational Linguistics. https://aclanthology.org/2025.acl-long.38/

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units involved in international projects. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Dominik Macko

Senior researcher, KInIT

Large language models (LLMs) are increasingly being used for a wide range of tasks and in various contexts in an interactive (chat-like) manner. The challenge is to make this interaction aligned with human expectations and values while avoiding reinforcing biases or negative social behavior of a model such as sycophancy, manipulativeness, etc. To do this, we need to be able to measure the extent of a positive or a negative behavior of the model and (ideally) have large datasets of human-LLM interactions together with human preferences that can be used to tune the models. For the latter, synthetic data generated by models trained on actual users’ data can sometimes be used instead.

The main goal is to improve the human-LLM interaction by using human inputs (such as their preferences) to fine-tune LLMs’ conversational capabilities while controlling for unwanted behavior (such as biases, sycophancy, manipulativeness, etc.). Alternatively, the topic can focus on measurement of the negative or positive social behavior of LLMs in various situations and contexts, or on improving safety of the models while taking the context and user preferences into account. 

The application domains are diverse, including (but not limited to) auditing of LLMs’ safety with respect to various user groups (including children and young adults), misuse of LLMs for disinformation generation, or credibility signals detection and assessment.

Relevant publications:

• Aneta Zugecova, Dominik Macko, Ivan Srba, Robert Moro, Jakub Kopál, Katarína Marcinčinová, and Matúš Mesarčík. 2025. Evaluation of LLM Vulnerabilities to Being Misused for Personalized Disinformation Generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 780–797, Vienna, Austria. Association for Computational Linguistics. https://aclanthology.org/2025.acl-long.38/ 
• Ivan Srba, Olesya Razuvayevskaya, João A. Leite, Robert Moro, Ipek Baris Schlicht, Sara Tonelli, Francisco Moreno García, Santiago Barrio Lottmann, Denis Teyssou, Valentin Porcellini, Carolina Scarton, Kalina Bontcheva, and Maria Bielikova. 2025. A Survey on Automatic Credibility Assessment Using Textual Credibility Signals in the Era of Large Language Models. ACM Trans. Intell. Syst. Technol. Just Accepted (September 2025). https://doi.org/10.1145/3770077 

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in collaboration with industrial partners or researchers from highly respected research units involved in international projects. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Róbert Móro

Senior researcher, KInIT

The proliferation of AI-driven recommendation systems on social media has intensified the need for methods that can anticipate how users engage with presented content. While current “digital twin” approaches simulate the behaviour of specific individuals, this PhD project is motivated by the need to predict the next interaction of user archetypes, i.e., personas defined by demographic and interest attributes. With the rapid evolution of generative AI models, including LLMs and multimodal systems, their capabilities to forecast user actions such as clicking, liking, skipping, or dwelling on social-media content remain unexplored.

The core objective of the PhD research is to investigate and compare AI-based approaches for predicting the next user interaction in social media environments, focusing on user archetypes rather than individuals. To do this, the work will explore generative modelling strategies (e.g., LLM-based predictors), multimodal encoders, and hybrid architectures capable of processing and reacting to streamed online content. A significant part of the thesis will address real-time content annotation, identifying which text/image/video attributes best predict behavioural responses, and how Parameter-Efficient Finetuning Techniques (PEFTs) can be used to adapt models rapidly and cost-effectively. The project will also analyse evaluation methodologies for sequential user-interaction forecasts, including simulation-based benchmarks that reflect realistic platform dynamics.

The domain of application will be algorithmic auditing, a research practice that evaluates how algorithmic systems operate and whether they comply with legal, ethical, or performance expectations. More accurate next-interaction prediction models can make these auditing approaches substantially more authentic, enabling simulated agents to behave in a more organic, human-like manner when interacting with recommender systems. Such improvements can strengthen audit validity, allow testing at scale, and produce a deeper understanding of the impact that recommender design has on different user archetypes.

Relevant publications:

• Ivan Srba, Branislav Pecher, Jakub Simko, Robert Moro, Maria Bielikova. 2025. Model-based Algorithmic Auditing of Social Media AI Algorithms. In Proceedings of Fourth European Workshop on Algorithmic Fairness, Vol. 294. 439-445. https://proceedings.mlr.press/v294/srba25a.html
• Matej Mosnar, Adam Skurla, Branislav Pecher, Matus Tibensky, Jan Jakubcik, Adrian Bindas, Peter Sakalik, and Ivan Srba. 2025. Revisiting Algorithmic Audits of TikTok: Poor Reproducibility and Short-term Validity of Findings. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’25). Association for Computing Machinery, New York, NY, USA, 3357–3366. https://doi.org/10.1145/3726302.3730293

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in collaboration with industrial partners or researchers from highly respected research units involved in international projects. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Ivan Srba

Senior researcher, KInIT

Current challenges that the energy sector faces today include maintaining stability in the energy grid, aggregation of flexibility, operation of energy communities, smart charging of electric vehicles, and efficient use of battery storage. Renewable resources are strongly dependent on weather, so weather prediction is also a focus of interest.

The ENVIRO team at KInIT is dedicated to the aforementioned tasks, whereas the proposed solutions are based on machine learning and artificial intelligence. We investigate and experiment with various methods such as taking into account physical laws in machine learning models (physics-informed machine learning), transfer or federative learning, as well as reinforcement learning, a currently popular approach used to optimize decision-making in a dynamic environment. Foundational models and their comparison with traditional approaches also offer an interesting direction of research. We solve tasks that process time series as well as image data using computer vision methods.

The aim of the PhD thesis will be to propose a solution to a selected energy or environmental problem (for example, flexibility aggregation or climate modeling/weather forecasting) by applying advanced artificial intelligence approaches. The selection of appropriate machine learning methods will depend on the specifics of the problem being solved and the choice of the doctoral candidate.

Relevant publications:

• Javed, H., Eid, F., El-Sappagh, S. et al. Sustainable energy management in the AI era: a comprehensive analysis of ML and DL approaches. Computing 107, 132 (2025). https://doi.org/10.1007/s00607-025-01485-0

• Pavlík, Peter, et al. „Fully differentiable Lagrangian convolutional neural network for physics-informed precipitation nowcasting.“ Applied Computing and Geosciences (2025): 100296. https://www.sciencedirect.com/science/article/pii/S2590197425000783

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Viera Rozinajová

Lead researcher, KInIT

Anna Bou Ezzeddine

Senior researcher, KInIT

Gabriela Grmanová

Senior researcher, KInIT

The rapid progress of large language models (LLMs) and other foundation models has transformed natural language processing (NLP), enabling powerful applications such as conversational agents, code assistants, and advanced information access systems. These technologies already affect everyday life and are reshaping how organisations work with text and knowledge.

Despite their success, LLMs still face important open challenges. Many models behave as black boxes, can hallucinate content, reproduce social biases, or fail when moved to new domains, tasks, or low-resource languages. At the same time, there is growing demand for models that are trustworthy, efficient, and easy to adapt, while respecting data, safety, and regulatory constraints.

This PhD project will investigate methods for understanding, adapting, and controlling large language models, with a particular focus on trustworthiness and low-resource / domain-specific NLP. The work can combine theory, modelling, and practical experimentation on real-world datasets.

Illustrative research challenges include:

• Properties of LLMs – analysing and mitigating hallucinations, robustness issues, and failure modes.
• Trustworthy NLP – bias detection and mitigation, safety and reliability, explainability and interpretability of model decisions.
• Adapting LLMs to specific contexts and tasks – e.g. parameter-efficient fine-tuning (PEFT), retrieval-augmented generation (RAG), instruction tuning, or continual learning.
• Advanced learning techniques – transfer and multilingual learning, active learning, low-shot and low-resource methods for under-represented languages.
• Domain-specific NLP applications – information extraction and text classification for specialised domains (e.g. legal, financial, technical or customer-support texts), improving dialog quality in conversational agents, or building task-oriented assistants.

The topic offers substantial flexibility and can be tailored to the candidate’s interests, ranging from more theoretical work on model behaviour to application-driven research in collaboration with academic or industrial partners.

Relevant publications:

• Vykopal, I., Ostermann, S., Simko, M. Soft Language Prompts for Language Transfer. In Proceedings of the 2025 Conference of the NAACL: Human Language Technologies (Volume 1: Long Papers), pages 10294-10313, ACL, 2025. https://doi.org/10.18653/v1/2025.naacl-long.517 

• Pikuliak, M., Hrčková, A., Oreško, Š., Šimko, M. Women Are Beautiful, Men Are Leaders: Gender Stereotypes in Machine Translation and Language Modeling. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3060-3083, ACL, 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.173   

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Marián Šimko

Lead researcher, KInIT

Jana Kosecka

Professor, George Mason University, USA

Large language models (LLMs) are currently driving much of the progress in the area of artificial intelligence, especially due to their ability to perform a wide range of tasks zero-shot – based on a natural language description of the task – and due to the large amount of world knowledge they possess.

However, the current generation of LLMs still struggles in several important ways that bar them from becoming truly capable at a wide range of challenging tasks; these struggles include e.g. their:

• Limited ability to correctly and robustly execute algorithms: Even frontier models currently struggle with correct execution of algorithms, especially when such execution is prolonged. Even models that theoretically support input contexts with millions of tokens, struggle to maintain accuracy as the context length grows.
• Limited ability to bootstrap complex reasoning: While models are now being explicitly trained to perform deliberative reasoning, multiple works demonstrate that the current generation of training methods used in this context (such as GRPO and GSPO) do not successfully bootstrap to new, improved reasoning skills – rather, they just reinforce reasoning paths that the base model already knew from pretraining.

Apart from these, there is also the limited ability to efficiently learn from experience (which is a component in how humans achieve true competence – we possess powerful priors, but we can also adapt quickly) and many other struggles. While this is – to a very considerable extent – true of all current models, the issues are typically much more pronounced for small models. This, of course, has serious implications for a lot of applications that require compact models, which can be run locally.

The aim of this research is to:

• Diagnose: Examine where the current models and training approaches struggle and why. In this phase, the student can draw upon existing insights (such as works on over-squashing and representational collapse in transformers), form their own hypotheses and validate them using the tools of mechanistic interpretability, use analogies with cases where similar approaches work (e.g. reinforcement learning methods like AlphaZero and MuZero – or even analogies from cognitive science regarding different types of human memory and such), etc.
• Fix: Use insights formed in the diagnosing phase to mitigate some of the issues – by adjusting the identified weaknesses, such as certain constraints imposed by the architecture, the training methodology, the training data and how it is preprocessed, etc.

Relevant publications:

• Gurgurov, Daniil, Michal Gregor, Josef van Genabith, Simon Ostermann. „On Multilingual Encoder Language Model Compression for Low-Resource Languages“. International Joint Conference on Natural Language Processing & Asia-Pacific Chapter of the Association for Computational Linguistics, 2025
• Vykopal, I., Pikuliak, M., Ostermann, S., Anikina, T., Gregor, M. and Šimko, M., 2025. Large Language Models for Multilingual Previously Fact-Checked Claim Detection. EMNLP 2025 (pp. 15741–15765).

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected.

Supervising team

Michal Gregor

Lead researcher, KInIT

Marián Šimko

Lead researcher, KInIT