Academically oriented

Combined study

4 years

prof. Ing. Lukáš Sekanina, Ph.D.

Chairman :
prof. Ing. Lukáš Sekanina, Ph.D.
Councillor internal :
doc. Ing. Jan Kořenek, Ph.D.
prof. Dr. Ing. Jan Černocký
prof. Ing. Adam Herout, Ph.D.
prof. Ing. Tomáš Hruška, CSc.
doc. Dr. Ing. Petr Hanáček
prof. Dr. Ing. Pavel Zemčík, dr. h. c.
prof. RNDr. Alexandr Meduna, CSc.
prof. Dr. Ing. Zbyněk Raida
prof. RNDr. Josef Šlapal, CSc.
prof. Ing. Pavel Václavek, Ph.D.
prof. Ing. Tomáš Vojnar, Ph.D.
Councillor external :
prof.,RNDr. Jiří Barnat, Ph.D.

The goal of the doctoral degree programme is to provide outstanding graduates from the master degree programme with a specialised university education of the highest level in certain fields of computer science and information technology, including especially the areas of information systems, computer-based systems and computer networks, computer graphics and multimedia, and intelligent systems. The education obtained within this degree programme also comprises a training and attestation for scientific work.

• Graduates from the doctoral study programme are trained to independently work in research, development, or management.
• They are able to solve and/or to lead teams solving advanced conceptual, research, development, or production problems in the area of contemporary information technology and its applications.
• They can be engaged to work on creative tasks, to lead research and development teams, or to work in management of companies or organizations whenever there are required abilities to work in an independent and creative way, to analyze complex problems, and to propose and realize new and original solutions. Graduates from the doctoral study programme can also teach and/or scientifically work at universities.

FIT graduates in general and FIT doctoral graduates in particular do not have a problem finding employment at scientific, pedagogical or management positions both in Czech Republic and abroad.

• FIT   graduates of the doctoral study are capable of independent scientific, research and management work in the field of Informatics, Computer Technology and Information Technologies. Graduates are ready to solve challenging conceptual, research and development problems. They can independently conduct research, development and production in the field of modern information technology.
• Typically, they work as creative workers at top scientific research workplaces, as leaders of research and development teams and in scientific and pedagogical work at universities. Graduates of this program are also employed in higher functional positions of larger institutions and companies, where the ability to work independently, analyze complex problems and design and implement new, original solutions is required.
• And, last but not least, graduates typically continue as so-called "postdoc" in their academic careers in Czech Republic or abroad.

The requirements that the doctoral students have to fulfil are given by their individual study plans, which specify the courses that they have to complete, their presupposed study visits and active participation at scientific conferences, and their minimum pedagogical activities within the bachelor and master degree programmes of the faculty. A successful completion of the doctoral studies is conditional on the following:

• The student has to pass a doctoral state examination within which he/she has to prove a deep knowledge of methodologies, theories, and their applications in accordance with the state of the art in the areas of science that are given by the courses included in his/her individual study plan and by the theme of his/her future dissertation thesis. The doctoral state examination also encompasses an evaluation of the presumed goals of the future dissertation thesis of the student, of the chosen solution method, and of the so far obtained original results.
• The student has further to prepare and defend his dissertation thesis.

The rules are determined by the directions of the dean for preparing the individual study plan of a doctoral student.  The plan is to be based on the theme of his/her future dissertation thesis and it is to be approved by the board of the branch.

• obligatory doctoral study programme Courses, the total number of courses a student has to complete and their mapping into particular semesters.
• a Research Plan Content (brief descrition of research content - focuse at the intended research area and the doctoral thesis topic
• a Research Plan ( list of research activities focused at the intended research area and the doctoral thesis topic - conferences and seminars to be attended , work to be published)
• teaching duty according to BUT study rules and regulations
• doctoral study schedule

Brno university of technology provides studies for persons with health disabilities according to section 21 par. 1 e) of the Act no. 111/1998, about universities and about the change and supplementing other laws (Higher Education Act) as amended, and according to the requirements in this field arising from Government Regulation No. 274/2016 Coll., on standards for accreditation in higher education, provides services for study applicants and students with specific needs within the scope and in form corresponding with the specification stated in Annex III to Rules for allocation of a financial contribution and funding for public universities by the Ministry of Education, Youth and Sports, specifying financing additional costs of studies for students with specific needs.

Services for students with specific needs at BUT are carried out through the activities of specialized workplace - Alfons counselling center, which is a part of BUT Lifelong Learning Institute - Student counselling section.

Counselling center activities and rules for making studies accessible are guaranteed by the university through a valid Rector's directive 11/2017 concerning the status of study applicants and students with specific needs at BUT. This internal standard guarantees minimal stadards of provided services.
Services of the counselling center are offered to all study applicants and students with any and all types of health disabilities stated in the Methodological standard of the Ministry of Education, Youth and Sports.

The study programme builds on both the ongoing follow-up Master's programme in Information Technology and the new follow-up Master's programme in Information Technology and Artificial Intelligence.
Students can also, according to their needs and outside their formalized studies, take courses and trainings related to the methodology of scientific work, publishing and citation skills, ethics, pedagogy and soft skills organized by BUT or other institutions.

3. round (applications submitted from 21.09.2026 to 31.01.2027)

1. Automated Techniques for Requirements-Based Test Assessment

   Evaluation of the completeness of test suites with respect to software requirements represents a fundamental challenge in the verification of critical systems, particularly in regulated domains. Current practice is largely based on manual assessment, which is time-consuming and error-prone. This dissertation will focus on the research of formal and automated methods for analyzing requirement coverage by test cases, with an emphasis on the semantic aspects of requirements and their implicit behavior.

   The goal of the dissertation is to propose a formal framework for representing requirements and test cases, and to define objective coverage criteria derived from the semantics of requirements (e.g., combinations of conditions, temporal constraints, sequences of actions, or state transitions). On this basis, methods for computing the coverage level of a test suite and for identifying uncovered semantic cases will be investigated.

   The research will focus in particular on the following areas:

   • design and systematic comparison of formal models suitable for representing different classes of requirements and testing artifacts, including analysis of their expressive power and analyzability;
   • formalization of the notion of requirement coverage and derivation of corresponding coverage obligations from the formal semantics of requirements;
   • design of algorithms for analyzing test suite coverage over formal representations, including identification of uncovered cases and their linkage to specific parts of the requirements;
   • investigation of methods for transforming informal requirements and test cases into formal models, including the use of natural language processing techniques and large language models, with emphasis on the reliability and limitations of these approaches;
   • experimental evaluation of the proposed methods on realistic datasets, including assessment of accuracy, completeness, and scalability.

   The dissertation will also include the design and implementation of prototype tools supporting selected aspects of the proposed approach, along with their validation on examples from embedded and avionics systems. Emphasis will be placed on integrating formal methods, software testing, and automated analysis techniques, with the aim of contributing to systematic and repeatable evaluation of test suite quality.

   Supervisor: Rogalewicz Adam, doc. Mgr., Ph.D.

2. Automated Techniques for Requirements-Based Test Assessment

   Evaluation of the completeness of test suites with respect to software requirements represents a fundamental challenge in the verification of critical systems, particularly in regulated domains. Current practice is largely based on manual assessment, which is time-consuming and error-prone. This dissertation will focus on the research of formal and automated methods for analyzing requirement coverage by test cases, with an emphasis on the semantic aspects of requirements and their implicit behavior.

   The goal of the dissertation is to propose a formal framework for representing requirements and test cases, and to define objective coverage criteria derived from the semantics of requirements (e.g., combinations of conditions, temporal constraints, sequences of actions, or state transitions). On this basis, methods for computing the coverage level of a test suite and for identifying uncovered semantic cases will be investigated.

   The research will focus in particular on the following areas:

   • design and systematic comparison of formal models suitable for representing different classes of requirements and testing artifacts, including analysis of their expressive power and analyzability;
   • formalization of the notion of requirement coverage and derivation of corresponding coverage obligations from the formal semantics of requirements;
   • design of algorithms for analyzing test suite coverage over formal representations, including identification of uncovered cases and their linkage to specific parts of the requirements;
   • investigation of methods for transforming informal requirements and test cases into formal models, including the use of natural language processing techniques and large language models, with emphasis on the reliability and limitations of these approaches;
   • experimental evaluation of the proposed methods on realistic datasets, including assessment of accuracy, completeness, and scalability.

   The dissertation will also include the design and implementation of prototype tools supporting selected aspects of the proposed approach, along with their validation on examples from embedded and avionics systems. Emphasis will be placed on integrating formal methods, software testing, and automated analysis techniques, with the aim of contributing to systematic and repeatable evaluation of test suite quality.

   Supervisor: Rogalewicz Adam, doc. Mgr., Ph.D.

3. Competent Reasoning and Reliable Execution – Towards Truly Capable LLMs

   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 ways that bar them from becoming truly capable at a wide range of challenging tasks; there is e.g. their:

   • Limited ability to correctly and robustly execute algorithms: Even frontier models struggle with correct execution of algorithms (especially when it is prolonged).
   • 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 (such as GRPO, GSPO, etc.) do not successfully bootstrap to new, improved reasoning skills, but mostly just reinforce reasoning paths that the model learned during pretraining.

   There is also the limited ability to efficiently learn from experience on the job (which is a component in how humans achieve true competence) and many other struggles. While this is – to a large extent – true of all current models, the issues are typically especially 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 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.

   Supervisor: Gregor Michal, doc. Ing., Ph.D.

4. Competent Reasoning and Reliable Execution – Towards Truly Capable LLMs

   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 ways that bar them from becoming truly capable at a wide range of challenging tasks; there is e.g. their:

   • Limited ability to correctly and robustly execute algorithms: Even frontier models struggle with correct execution of algorithms (especially when it is prolonged).
   • 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 (such as GRPO, GSPO, etc.) do not successfully bootstrap to new, improved reasoning skills, but mostly just reinforce reasoning paths that the model learned during pretraining.

   There is also the limited ability to efficiently learn from experience on the job (which is a component in how humans achieve true competence) and many other struggles. While this is – to a large extent – true of all current models, the issues are typically especially 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 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.

   Supervisor: Gregor Michal, doc. Ing., Ph.D.

5. Efficient Technology for Dealing with Logics and Automata (Not Only) for Formal Analysis and Verification

   Different types of logics and automata are among the most fundamental objects studied and applied in computer science for decades. Nevertheless, there are many unsatisfactorily solved problems in this area, and new, exciting problems related to ever new applications of logics and automata are constantly emerging (e.g., in the formal verification of finite and infinite state systems with various advanced control or data structures, in decision procedures, in program or hardware synthesis, in quantum program verification, or even in methods for efficient search in various types of data or network traffic).

   The subject of the dissertation will primarily be the development of the state of the art in efficient work with various logics (e.g. over pointer structures, strings, various arithmetic, temporal logics, etc.). To this end, approaches based on different types of decision diagrams, automata, but also e.g. approaches based on the existence of a model of bounded size or on efficient reductions between different types of logical theories will be investigated. In connection with this, methods for efficient work with decision diagrams and different types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will be developed. The work will include theoretical research as well as prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be solved in collaboration with the VeriFIT team working on the development of techniques for working with logics and automata and their applications at FIT BUT. There is also the possibility of close cooperation with various foreign partners of VeriFIT: Academia Sinica, Taiwan (Prof. Y. Chen); TU Vienna, Austria (Assoc. F. Zuleger); LSV, ENS Paris-Saclay (Assoc. M. Sighireanu); IRIF, Paris, France (Assoc. A. Bouajjani, Assoc. P. Habermehl, Assoc. C. Enea), Verimag, Grenoble, France (Assoc. R. Iosif); Uppsala University, Sweden (prof. P.A. Abdulla, prof. B. Jonsson); or RPTU, Kaiserslautern, Germany (prof. A.W. Lin).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

6. Efficient Technology for Dealing with Logics and Automata (Not Only) for Formal Analysis and Verification

   Different types of logics and automata are among the most fundamental objects studied and applied in computer science for decades. Nevertheless, there are many unsatisfactorily solved problems in this area, and new, exciting problems related to ever new applications of logics and automata are constantly emerging (e.g., in the formal verification of finite and infinite state systems with various advanced control or data structures, in decision procedures, in program or hardware synthesis, in quantum program verification, or even in methods for efficient search in various types of data or network traffic).

   The subject of the dissertation will primarily be the development of the state of the art in efficient work with various logics (e.g. over pointer structures, strings, various arithmetic, temporal logics, etc.). To this end, approaches based on different types of decision diagrams, automata, but also e.g. approaches based on the existence of a model of bounded size or on efficient reductions between different types of logical theories will be investigated. In connection with this, methods for efficient work with decision diagrams and different types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will be developed. The work will include theoretical research as well as prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be solved in collaboration with the VeriFIT team working on the development of techniques for working with logics and automata and their applications at FIT BUT. There is also the possibility of close cooperation with various foreign partners of VeriFIT: Academia Sinica, Taiwan (Prof. Y. Chen); TU Vienna, Austria (Assoc. F. Zuleger); LSV, ENS Paris-Saclay (Assoc. M. Sighireanu); IRIF, Paris, France (Assoc. A. Bouajjani, Assoc. P. Habermehl, Assoc. C. Enea), Verimag, Grenoble, France (Assoc. R. Iosif); Uppsala University, Sweden (prof. P.A. Abdulla, prof. B. Jonsson); or RPTU, Kaiserslautern, Germany (prof. A.W. Lin).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

7. Formal Methods in Design and Analysis of Quantum Programs

   Development in the field of quantum computers is moving forward unstoppably. However, for the effective use of quantum hardware, it is necessary to have the right support at the software level, i.e., quantum algorithms for solving problems, as well as tools for compiling, optimizing and synthesizing quantum programs, mapping them to a specific quantum technology, mapping them to a fault tolerant protocol, analyzing, simulating, and debugging quantum programs, etc.

   The subject of the dissertation will primarily be the development of current knowledge in the field of formal methods in quantum program design. The main focus here will be on the use of automata theory, automatic reasoning, and decision diagrams. In connection with this, methods for effective work with decision diagrams and various types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will also be developed. The work will include both theoretical research and prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be carried out in collaboration with the VeriFIT team at FIT BUT, which develops techniques for working with logics and automata and their applications. There is also the possibility of close cooperation with various foreign VeriFIT partners: Academia Sinica, Taiwan (prof. Y. Chen); Uppsala University, Sweden (prof. P.A. Abdulla, prof. R. S. Thinniyam); or Leiden Institute for Advanced Computer Science, Leiden, the Netherlands (prof. A. Laarman).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

8. Formal Methods in Design and Analysis of Quantum Programs

   Development in the field of quantum computers is moving forward unstoppably. However, for the effective use of quantum hardware, it is necessary to have the right support at the software level, i.e., quantum algorithms for solving problems, as well as tools for compiling, optimizing and synthesizing quantum programs, mapping them to a specific quantum technology, mapping them to a fault tolerant protocol, analyzing, simulating, and debugging quantum programs, etc.

   The subject of the dissertation will primarily be the development of current knowledge in the field of formal methods in quantum program design. The main focus here will be on the use of automata theory, automatic reasoning, and decision diagrams. In connection with this, methods for effective work with decision diagrams and various types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will also be developed. The work will include both theoretical research and prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be carried out in collaboration with the VeriFIT team at FIT BUT, which develops techniques for working with logics and automata and their applications. There is also the possibility of close cooperation with various foreign VeriFIT partners: Academia Sinica, Taiwan (prof. Y. Chen); Uppsala University, Sweden (prof. P.A. Abdulla, prof. R. S. Thinniyam); or Leiden Institute for Advanced Computer Science, Leiden, the Netherlands (prof. A. Laarman).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

9. Human in the loop (HIL) in the adaptation of AI systems

   This PhD topic is driven by the needs of users for which the privacy of data is crucial and concentrates on improving AI (especially speech recognition) system in case the user is able to identify and correct systems errors. The topic includes proper evaluations of HIL systems, selection of data to be proposed for correction and actual fine-tuning / adaptation techniques working with large pre-trained models. 

   Supervisor: Černocký Jan, prof. Dr. Ing.

10. Human in the loop (HIL) in the adaptation of AI systems

    This PhD topic is driven by the needs of users for which the privacy of data is crucial and concentrates on improving AI (especially speech recognition) system in case the user is able to identify and correct systems errors. The topic includes proper evaluations of HIL systems, selection of data to be proposed for correction and actual fine-tuning / adaptation techniques working with large pre-trained models. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

11. Improving human-LLM interaction

    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.

    Supervisor: Móro Róbert, Ing., Ph.D.

12. Improving human-LLM interaction

    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.

    Supervisor: Móro Róbert, Ing., Ph.D.

13. Improving natural language processing

    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. 

    Supervisor: Šimko Marián, doc. Ing., Ph.D.

14. Improving natural language processing

    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. 

    Supervisor: Šimko Marián, doc. Ing., Ph.D.

15. Low-resource text processing with the help of data synthesis

    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.

    Supervisor: Šimko Jakub, doc. Ing., PhD.

16. Low-resource text processing with the help of data synthesis

    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.

    Supervisor: Šimko Jakub, doc. Ing., PhD.

17. Meta-evaluation of multilingual texts by Large Language Models

    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.

    Supervisor: Macko Dominik, doc. Ing., Ph.D.

18. Meta-evaluation of multilingual texts by Large Language Models

    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.

    Supervisor: Macko Dominik, doc. Ing., Ph.D.

19. Multimodal analysis for assessment of mental health

    Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.

    Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.

    Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.

    Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.

    Few Words About Supervision: I have extensive experience of working in the field of neuro-signal and neuroimage processing and I am currently head of a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz

    Supervisor: Malik Aamir Saeed, prof., Ph.D.

20. Multimodal analysis for assessment of mental health

    Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.

    Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.

    Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.

    Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.

    Few Words About Supervision: I have extensive experience of working in the field of neuro-signal and neuroimage processing and I am currently head of a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz

    Supervisor: Malik Aamir Saeed, prof., Ph.D.

21. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

22. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

23. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

24. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

25. Recommender and adaptive web-based systems

    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.

    Supervisor: Kompan Michal, doc. Ing., PhD.

26. Recommender and adaptive web-based systems

    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.

    Supervisor: Kompan Michal, doc. Ing., PhD.

27. Research and advancement of the visualization, interaction and AI technique Malleable Glyph

    The goal of the doctoral program will be to develop the newly proposed technique, Malleable Glyphs, in the fields of human-computer interaction, visualization, and computer vision. The researcher will develop machine learning and computer vision algorithms for glyphs. The researcher will also co-organize The Malleable Glyph Challenge, design new glyph modalities, evaluate glyphs submitted by students and individuals from third-party institutions, and create a glyph taxonomy.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

28. Research and advancement of the visualization, interaction and AI technique Malleable Glyph

    The goal of the doctoral program will be to develop the newly proposed technique, Malleable Glyphs, in the fields of human-computer interaction, visualization, and computer vision. The researcher will develop machine learning and computer vision algorithms for glyphs. The researcher will also co-organize The Malleable Glyph Challenge, design new glyph modalities, evaluate glyphs submitted by students and individuals from third-party institutions, and create a glyph taxonomy.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

29. Speech recognition and data processing for ATC-pilot communication in aviation

    This PhD topic is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology), language model as well as interaction with sources of meta-data (radar information, airport). Special attention will be paid to the use of large pre-trained model, code-switching, reliable language identification from short segments and adaptation to the conditions of given airport with crawled data.

    Supervisor: Černocký Jan, prof. Dr. Ing.

30. Speech recognition and data processing for ATC-pilot communication in aviation

    This PhD topic is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology), language model as well as interaction with sources of meta-data (radar information, airport). Special attention will be paid to the use of large pre-trained model, code-switching, reliable language identification from short segments and adaptation to the conditions of given airport with crawled data.

    Supervisor: Černocký Jan, prof. Dr. Ing.

31. User Activity Prediction in Social Media Environments

    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.
    • 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 SIGIR '25. ACM, New York, NY, USA, 3357–3366.

    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.

    Supervisor: Srba Ivan, Ing., Ph.D.

32. User Activity Prediction in Social Media Environments

    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.
    • 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 SIGIR '25. ACM, New York, NY, USA, 3357–3366.

    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.

    Supervisor: Srba Ivan, Ing., Ph.D.

33. User Experience in Modern User Interfaces

    The goal of the doctoral study is to analyze the characteristics of modern and popular user interfaces, with a focus on cognitive load, addiction-inducing techniques, the perception of AI-generated content, the perception of advertising, and other significant contemporary phenomena. The student will prototype user interface elements and design and conduct user tests that will enable a deeper understanding of the phenomena under study, highlight existing risks, and lead to the design of user interface elements that will represent alternatives beneficial to the user.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

34. User Experience in Modern User Interfaces

    The goal of the doctoral study is to analyze the characteristics of modern and popular user interfaces, with a focus on cognitive load, addiction-inducing techniques, the perception of AI-generated content, the perception of advertising, and other significant contemporary phenomena. The student will prototype user interface elements and design and conduct user tests that will enable a deeper understanding of the phenomena under study, highlight existing risks, and lead to the design of user interface elements that will represent alternatives beneficial to the user.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

35. Using machine learning to solve energy and environmental problems

    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

    Supervisor: Rozinajová Věra, Doc., Ph.D.

36. Using machine learning to solve energy and environmental problems

    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

    Supervisor: Rozinajová Věra, Doc., Ph.D.

2. round (applications submitted from 01.06.2026 to 30.08.2026)

1. Automated Techniques for Requirements-Based Test Assessment

   Evaluation of the completeness of test suites with respect to software requirements represents a fundamental challenge in the verification of critical systems, particularly in regulated domains. Current practice is largely based on manual assessment, which is time-consuming and error-prone. This dissertation will focus on the research of formal and automated methods for analyzing requirement coverage by test cases, with an emphasis on the semantic aspects of requirements and their implicit behavior.

   The goal of the dissertation is to propose a formal framework for representing requirements and test cases, and to define objective coverage criteria derived from the semantics of requirements (e.g., combinations of conditions, temporal constraints, sequences of actions, or state transitions). On this basis, methods for computing the coverage level of a test suite and for identifying uncovered semantic cases will be investigated.

   The research will focus in particular on the following areas:

   • design and systematic comparison of formal models suitable for representing different classes of requirements and testing artifacts, including analysis of their expressive power and analyzability;
   • formalization of the notion of requirement coverage and derivation of corresponding coverage obligations from the formal semantics of requirements;
   • design of algorithms for analyzing test suite coverage over formal representations, including identification of uncovered cases and their linkage to specific parts of the requirements;
   • investigation of methods for transforming informal requirements and test cases into formal models, including the use of natural language processing techniques and large language models, with emphasis on the reliability and limitations of these approaches;
   • experimental evaluation of the proposed methods on realistic datasets, including assessment of accuracy, completeness, and scalability.

   The dissertation will also include the design and implementation of prototype tools supporting selected aspects of the proposed approach, along with their validation on examples from embedded and avionics systems. Emphasis will be placed on integrating formal methods, software testing, and automated analysis techniques, with the aim of contributing to systematic and repeatable evaluation of test suite quality.

   Supervisor: Rogalewicz Adam, doc. Mgr., Ph.D.

2. Automated Techniques for Requirements-Based Test Assessment

   Evaluation of the completeness of test suites with respect to software requirements represents a fundamental challenge in the verification of critical systems, particularly in regulated domains. Current practice is largely based on manual assessment, which is time-consuming and error-prone. This dissertation will focus on the research of formal and automated methods for analyzing requirement coverage by test cases, with an emphasis on the semantic aspects of requirements and their implicit behavior.

   The goal of the dissertation is to propose a formal framework for representing requirements and test cases, and to define objective coverage criteria derived from the semantics of requirements (e.g., combinations of conditions, temporal constraints, sequences of actions, or state transitions). On this basis, methods for computing the coverage level of a test suite and for identifying uncovered semantic cases will be investigated.

   The research will focus in particular on the following areas:

   • design and systematic comparison of formal models suitable for representing different classes of requirements and testing artifacts, including analysis of their expressive power and analyzability;
   • formalization of the notion of requirement coverage and derivation of corresponding coverage obligations from the formal semantics of requirements;
   • design of algorithms for analyzing test suite coverage over formal representations, including identification of uncovered cases and their linkage to specific parts of the requirements;
   • investigation of methods for transforming informal requirements and test cases into formal models, including the use of natural language processing techniques and large language models, with emphasis on the reliability and limitations of these approaches;
   • experimental evaluation of the proposed methods on realistic datasets, including assessment of accuracy, completeness, and scalability.

   The dissertation will also include the design and implementation of prototype tools supporting selected aspects of the proposed approach, along with their validation on examples from embedded and avionics systems. Emphasis will be placed on integrating formal methods, software testing, and automated analysis techniques, with the aim of contributing to systematic and repeatable evaluation of test suite quality.

   Supervisor: Rogalewicz Adam, doc. Mgr., Ph.D.

3. Competent Reasoning and Reliable Execution – Towards Truly Capable LLMs

   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 ways that bar them from becoming truly capable at a wide range of challenging tasks; there is e.g. their:

   • Limited ability to correctly and robustly execute algorithms: Even frontier models struggle with correct execution of algorithms (especially when it is prolonged).
   • 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 (such as GRPO, GSPO, etc.) do not successfully bootstrap to new, improved reasoning skills, but mostly just reinforce reasoning paths that the model learned during pretraining.

   There is also the limited ability to efficiently learn from experience on the job (which is a component in how humans achieve true competence) and many other struggles. While this is – to a large extent – true of all current models, the issues are typically especially 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 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.

   Supervisor: Gregor Michal, doc. Ing., Ph.D.

4. Competent Reasoning and Reliable Execution – Towards Truly Capable LLMs

   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 ways that bar them from becoming truly capable at a wide range of challenging tasks; there is e.g. their:

   • Limited ability to correctly and robustly execute algorithms: Even frontier models struggle with correct execution of algorithms (especially when it is prolonged).
   • 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 (such as GRPO, GSPO, etc.) do not successfully bootstrap to new, improved reasoning skills, but mostly just reinforce reasoning paths that the model learned during pretraining.

   There is also the limited ability to efficiently learn from experience on the job (which is a component in how humans achieve true competence) and many other struggles. While this is – to a large extent – true of all current models, the issues are typically especially 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 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.

   Supervisor: Gregor Michal, doc. Ing., Ph.D.

5. Efficient Technology for Dealing with Logics and Automata (Not Only) for Formal Analysis and Verification

   Different types of logics and automata are among the most fundamental objects studied and applied in computer science for decades. Nevertheless, there are many unsatisfactorily solved problems in this area, and new, exciting problems related to ever new applications of logics and automata are constantly emerging (e.g., in the formal verification of finite and infinite state systems with various advanced control or data structures, in decision procedures, in program or hardware synthesis, in quantum program verification, or even in methods for efficient search in various types of data or network traffic).

   The subject of the dissertation will primarily be the development of the state of the art in efficient work with various logics (e.g. over pointer structures, strings, various arithmetic, temporal logics, etc.). To this end, approaches based on different types of decision diagrams, automata, but also e.g. approaches based on the existence of a model of bounded size or on efficient reductions between different types of logical theories will be investigated. In connection with this, methods for efficient work with decision diagrams and different types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will be developed. The work will include theoretical research as well as prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be solved in collaboration with the VeriFIT team working on the development of techniques for working with logics and automata and their applications at FIT BUT. There is also the possibility of close cooperation with various foreign partners of VeriFIT: Academia Sinica, Taiwan (Prof. Y. Chen); TU Vienna, Austria (Assoc. F. Zuleger); LSV, ENS Paris-Saclay (Assoc. M. Sighireanu); IRIF, Paris, France (Assoc. A. Bouajjani, Assoc. P. Habermehl, Assoc. C. Enea), Verimag, Grenoble, France (Assoc. R. Iosif); Uppsala University, Sweden (prof. P.A. Abdulla, prof. B. Jonsson); or RPTU, Kaiserslautern, Germany (prof. A.W. Lin).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

6. Efficient Technology for Dealing with Logics and Automata (Not Only) for Formal Analysis and Verification

   Different types of logics and automata are among the most fundamental objects studied and applied in computer science for decades. Nevertheless, there are many unsatisfactorily solved problems in this area, and new, exciting problems related to ever new applications of logics and automata are constantly emerging (e.g., in the formal verification of finite and infinite state systems with various advanced control or data structures, in decision procedures, in program or hardware synthesis, in quantum program verification, or even in methods for efficient search in various types of data or network traffic).

   The subject of the dissertation will primarily be the development of the state of the art in efficient work with various logics (e.g. over pointer structures, strings, various arithmetic, temporal logics, etc.). To this end, approaches based on different types of decision diagrams, automata, but also e.g. approaches based on the existence of a model of bounded size or on efficient reductions between different types of logical theories will be investigated. In connection with this, methods for efficient work with decision diagrams and different types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will be developed. The work will include theoretical research as well as prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be solved in collaboration with the VeriFIT team working on the development of techniques for working with logics and automata and their applications at FIT BUT. There is also the possibility of close cooperation with various foreign partners of VeriFIT: Academia Sinica, Taiwan (Prof. Y. Chen); TU Vienna, Austria (Assoc. F. Zuleger); LSV, ENS Paris-Saclay (Assoc. M. Sighireanu); IRIF, Paris, France (Assoc. A. Bouajjani, Assoc. P. Habermehl, Assoc. C. Enea), Verimag, Grenoble, France (Assoc. R. Iosif); Uppsala University, Sweden (prof. P.A. Abdulla, prof. B. Jonsson); or RPTU, Kaiserslautern, Germany (prof. A.W. Lin).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

7. Formal Methods in Design and Analysis of Quantum Programs

   Development in the field of quantum computers is moving forward unstoppably. However, for the effective use of quantum hardware, it is necessary to have the right support at the software level, i.e., quantum algorithms for solving problems, as well as tools for compiling, optimizing and synthesizing quantum programs, mapping them to a specific quantum technology, mapping them to a fault tolerant protocol, analyzing, simulating, and debugging quantum programs, etc.

   The subject of the dissertation will primarily be the development of current knowledge in the field of formal methods in quantum program design. The main focus here will be on the use of automata theory, automatic reasoning, and decision diagrams. In connection with this, methods for effective work with decision diagrams and various types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will also be developed. The work will include both theoretical research and prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be carried out in collaboration with the VeriFIT team at FIT BUT, which develops techniques for working with logics and automata and their applications. There is also the possibility of close cooperation with various foreign VeriFIT partners: Academia Sinica, Taiwan (prof. Y. Chen); Uppsala University, Sweden (prof. P.A. Abdulla, prof. R. S. Thinniyam); or Leiden Institute for Advanced Computer Science, Leiden, the Netherlands (prof. A. Laarman).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

8. Formal Methods in Design and Analysis of Quantum Programs

   Development in the field of quantum computers is moving forward unstoppably. However, for the effective use of quantum hardware, it is necessary to have the right support at the software level, i.e., quantum algorithms for solving problems, as well as tools for compiling, optimizing and synthesizing quantum programs, mapping them to a specific quantum technology, mapping them to a fault tolerant protocol, analyzing, simulating, and debugging quantum programs, etc.

   The subject of the dissertation will primarily be the development of current knowledge in the field of formal methods in quantum program design. The main focus here will be on the use of automata theory, automatic reasoning, and decision diagrams. In connection with this, methods for effective work with decision diagrams and various types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will also be developed. The work will include both theoretical research and prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be carried out in collaboration with the VeriFIT team at FIT BUT, which develops techniques for working with logics and automata and their applications. There is also the possibility of close cooperation with various foreign VeriFIT partners: Academia Sinica, Taiwan (prof. Y. Chen); Uppsala University, Sweden (prof. P.A. Abdulla, prof. R. S. Thinniyam); or Leiden Institute for Advanced Computer Science, Leiden, the Netherlands (prof. A. Laarman).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

9. Human in the loop (HIL) in the adaptation of AI systems

   This PhD topic is driven by the needs of users for which the privacy of data is crucial and concentrates on improving AI (especially speech recognition) system in case the user is able to identify and correct systems errors. The topic includes proper evaluations of HIL systems, selection of data to be proposed for correction and actual fine-tuning / adaptation techniques working with large pre-trained models. 

   Supervisor: Černocký Jan, prof. Dr. Ing.

10. Human in the loop (HIL) in the adaptation of AI systems

    This PhD topic is driven by the needs of users for which the privacy of data is crucial and concentrates on improving AI (especially speech recognition) system in case the user is able to identify and correct systems errors. The topic includes proper evaluations of HIL systems, selection of data to be proposed for correction and actual fine-tuning / adaptation techniques working with large pre-trained models. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

11. Improving human-LLM interaction

    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.

    Supervisor: Móro Róbert, Ing., Ph.D.

12. Improving human-LLM interaction

    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.

    Supervisor: Móro Róbert, Ing., Ph.D.

13. Improving natural language processing

    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. 

    Supervisor: Šimko Marián, doc. Ing., Ph.D.

14. Improving natural language processing

    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. 

    Supervisor: Šimko Marián, doc. Ing., Ph.D.

15. Low-resource text processing with the help of data synthesis

    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.

    Supervisor: Šimko Jakub, doc. Ing., PhD.

16. Low-resource text processing with the help of data synthesis

    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.

    Supervisor: Šimko Jakub, doc. Ing., PhD.

17. Meta-evaluation of multilingual texts by Large Language Models

    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.

    Supervisor: Macko Dominik, doc. Ing., Ph.D.

18. Meta-evaluation of multilingual texts by Large Language Models

    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.

    Supervisor: Macko Dominik, doc. Ing., Ph.D.

19. Multimodal analysis for assessment of mental health

    Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.

    Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.

    Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.

    Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.

    Few Words About Supervision: I have extensive experience of working in the field of neuro-signal and neuroimage processing and I am currently head of a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz

    Supervisor: Malik Aamir Saeed, prof., Ph.D.

20. Multimodal analysis for assessment of mental health

    Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.

    Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.

    Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.

    Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.

    Few Words About Supervision: I have extensive experience of working in the field of neuro-signal and neuroimage processing and I am currently head of a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz

    Supervisor: Malik Aamir Saeed, prof., Ph.D.

21. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

22. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

23. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

24. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

25. Recommender and adaptive web-based systems

    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.

    Supervisor: Kompan Michal, doc. Ing., PhD.

26. Recommender and adaptive web-based systems

    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.

    Supervisor: Kompan Michal, doc. Ing., PhD.

27. Research and advancement of the visualization, interaction and AI technique Malleable Glyph

    The goal of the doctoral program will be to develop the newly proposed technique, Malleable Glyphs, in the fields of human-computer interaction, visualization, and computer vision. The researcher will develop machine learning and computer vision algorithms for glyphs. The researcher will also co-organize The Malleable Glyph Challenge, design new glyph modalities, evaluate glyphs submitted by students and individuals from third-party institutions, and create a glyph taxonomy.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

28. Research and advancement of the visualization, interaction and AI technique Malleable Glyph

    The goal of the doctoral program will be to develop the newly proposed technique, Malleable Glyphs, in the fields of human-computer interaction, visualization, and computer vision. The researcher will develop machine learning and computer vision algorithms for glyphs. The researcher will also co-organize The Malleable Glyph Challenge, design new glyph modalities, evaluate glyphs submitted by students and individuals from third-party institutions, and create a glyph taxonomy.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

29. Speech recognition and data processing for ATC-pilot communication in aviation

    This PhD topic is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology), language model as well as interaction with sources of meta-data (radar information, airport). Special attention will be paid to the use of large pre-trained model, code-switching, reliable language identification from short segments and adaptation to the conditions of given airport with crawled data.

    Supervisor: Černocký Jan, prof. Dr. Ing.

30. Speech recognition and data processing for ATC-pilot communication in aviation

    This PhD topic is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology), language model as well as interaction with sources of meta-data (radar information, airport). Special attention will be paid to the use of large pre-trained model, code-switching, reliable language identification from short segments and adaptation to the conditions of given airport with crawled data.

    Supervisor: Černocký Jan, prof. Dr. Ing.

31. User Activity Prediction in Social Media Environments

    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.
    • 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 SIGIR '25. ACM, New York, NY, USA, 3357–3366.

    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.

    Supervisor: Srba Ivan, Ing., Ph.D.

32. User Activity Prediction in Social Media Environments

    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.
    • 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 SIGIR '25. ACM, New York, NY, USA, 3357–3366.

    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.

    Supervisor: Srba Ivan, Ing., Ph.D.

33. User Experience in Modern User Interfaces

    The goal of the doctoral study is to analyze the characteristics of modern and popular user interfaces, with a focus on cognitive load, addiction-inducing techniques, the perception of AI-generated content, the perception of advertising, and other significant contemporary phenomena. The student will prototype user interface elements and design and conduct user tests that will enable a deeper understanding of the phenomena under study, highlight existing risks, and lead to the design of user interface elements that will represent alternatives beneficial to the user.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

34. User Experience in Modern User Interfaces

    The goal of the doctoral study is to analyze the characteristics of modern and popular user interfaces, with a focus on cognitive load, addiction-inducing techniques, the perception of AI-generated content, the perception of advertising, and other significant contemporary phenomena. The student will prototype user interface elements and design and conduct user tests that will enable a deeper understanding of the phenomena under study, highlight existing risks, and lead to the design of user interface elements that will represent alternatives beneficial to the user.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

35. Using machine learning to solve energy and environmental problems

    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

    Supervisor: Rozinajová Věra, Doc., Ph.D.

36. Using machine learning to solve energy and environmental problems

    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

    Supervisor: Rozinajová Věra, Doc., Ph.D.

1. round (applications submitted from 01.03.2026 to 31.05.2026)

1. Automated Techniques for Requirements-Based Test Assessment

   Evaluation of the completeness of test suites with respect to software requirements represents a fundamental challenge in the verification of critical systems, particularly in regulated domains. Current practice is largely based on manual assessment, which is time-consuming and error-prone. This dissertation will focus on the research of formal and automated methods for analyzing requirement coverage by test cases, with an emphasis on the semantic aspects of requirements and their implicit behavior.

   The goal of the dissertation is to propose a formal framework for representing requirements and test cases, and to define objective coverage criteria derived from the semantics of requirements (e.g., combinations of conditions, temporal constraints, sequences of actions, or state transitions). On this basis, methods for computing the coverage level of a test suite and for identifying uncovered semantic cases will be investigated.

   The research will focus in particular on the following areas:

   • design and systematic comparison of formal models suitable for representing different classes of requirements and testing artifacts, including analysis of their expressive power and analyzability;
   • formalization of the notion of requirement coverage and derivation of corresponding coverage obligations from the formal semantics of requirements;
   • design of algorithms for analyzing test suite coverage over formal representations, including identification of uncovered cases and their linkage to specific parts of the requirements;
   • investigation of methods for transforming informal requirements and test cases into formal models, including the use of natural language processing techniques and large language models, with emphasis on the reliability and limitations of these approaches;
   • experimental evaluation of the proposed methods on realistic datasets, including assessment of accuracy, completeness, and scalability.

   The dissertation will also include the design and implementation of prototype tools supporting selected aspects of the proposed approach, along with their validation on examples from embedded and avionics systems. Emphasis will be placed on integrating formal methods, software testing, and automated analysis techniques, with the aim of contributing to systematic and repeatable evaluation of test suite quality.

   Supervisor: Rogalewicz Adam, doc. Mgr., Ph.D.

2. Automated Techniques for Requirements-Based Test Assessment

   Evaluation of the completeness of test suites with respect to software requirements represents a fundamental challenge in the verification of critical systems, particularly in regulated domains. Current practice is largely based on manual assessment, which is time-consuming and error-prone. This dissertation will focus on the research of formal and automated methods for analyzing requirement coverage by test cases, with an emphasis on the semantic aspects of requirements and their implicit behavior.

   The goal of the dissertation is to propose a formal framework for representing requirements and test cases, and to define objective coverage criteria derived from the semantics of requirements (e.g., combinations of conditions, temporal constraints, sequences of actions, or state transitions). On this basis, methods for computing the coverage level of a test suite and for identifying uncovered semantic cases will be investigated.

   The research will focus in particular on the following areas:

   • design and systematic comparison of formal models suitable for representing different classes of requirements and testing artifacts, including analysis of their expressive power and analyzability;
   • formalization of the notion of requirement coverage and derivation of corresponding coverage obligations from the formal semantics of requirements;
   • design of algorithms for analyzing test suite coverage over formal representations, including identification of uncovered cases and their linkage to specific parts of the requirements;
   • investigation of methods for transforming informal requirements and test cases into formal models, including the use of natural language processing techniques and large language models, with emphasis on the reliability and limitations of these approaches;
   • experimental evaluation of the proposed methods on realistic datasets, including assessment of accuracy, completeness, and scalability.

   The dissertation will also include the design and implementation of prototype tools supporting selected aspects of the proposed approach, along with their validation on examples from embedded and avionics systems. Emphasis will be placed on integrating formal methods, software testing, and automated analysis techniques, with the aim of contributing to systematic and repeatable evaluation of test suite quality.

   Supervisor: Rogalewicz Adam, doc. Mgr., Ph.D.

3. Competent Reasoning and Reliable Execution – Towards Truly Capable LLMs

   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 ways that bar them from becoming truly capable at a wide range of challenging tasks; there is e.g. their:

   • Limited ability to correctly and robustly execute algorithms: Even frontier models struggle with correct execution of algorithms (especially when it is prolonged).
   • 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 (such as GRPO, GSPO, etc.) do not successfully bootstrap to new, improved reasoning skills, but mostly just reinforce reasoning paths that the model learned during pretraining.

   There is also the limited ability to efficiently learn from experience on the job (which is a component in how humans achieve true competence) and many other struggles. While this is – to a large extent – true of all current models, the issues are typically especially 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 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.

   Supervisor: Gregor Michal, doc. Ing., Ph.D.

4. Competent Reasoning and Reliable Execution – Towards Truly Capable LLMs

   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 ways that bar them from becoming truly capable at a wide range of challenging tasks; there is e.g. their:

   • Limited ability to correctly and robustly execute algorithms: Even frontier models struggle with correct execution of algorithms (especially when it is prolonged).
   • 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 (such as GRPO, GSPO, etc.) do not successfully bootstrap to new, improved reasoning skills, but mostly just reinforce reasoning paths that the model learned during pretraining.

   There is also the limited ability to efficiently learn from experience on the job (which is a component in how humans achieve true competence) and many other struggles. While this is – to a large extent – true of all current models, the issues are typically especially 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 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.

   Supervisor: Gregor Michal, doc. Ing., Ph.D.

5. Efficient Technology for Dealing with Logics and Automata (Not Only) for Formal Analysis and Verification

   Different types of logics and automata are among the most fundamental objects studied and applied in computer science for decades. Nevertheless, there are many unsatisfactorily solved problems in this area, and new, exciting problems related to ever new applications of logics and automata are constantly emerging (e.g., in the formal verification of finite and infinite state systems with various advanced control or data structures, in decision procedures, in program or hardware synthesis, in quantum program verification, or even in methods for efficient search in various types of data or network traffic).

   The subject of the dissertation will primarily be the development of the state of the art in efficient work with various logics (e.g. over pointer structures, strings, various arithmetic, temporal logics, etc.). To this end, approaches based on different types of decision diagrams, automata, but also e.g. approaches based on the existence of a model of bounded size or on efficient reductions between different types of logical theories will be investigated. In connection with this, methods for efficient work with decision diagrams and different types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will be developed. The work will include theoretical research as well as prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be solved in collaboration with the VeriFIT team working on the development of techniques for working with logics and automata and their applications at FIT BUT. There is also the possibility of close cooperation with various foreign partners of VeriFIT: Academia Sinica, Taiwan (Prof. Y. Chen); TU Vienna, Austria (Assoc. F. Zuleger); LSV, ENS Paris-Saclay (Assoc. M. Sighireanu); IRIF, Paris, France (Assoc. A. Bouajjani, Assoc. P. Habermehl, Assoc. C. Enea), Verimag, Grenoble, France (Assoc. R. Iosif); Uppsala University, Sweden (prof. P.A. Abdulla, prof. B. Jonsson); or RPTU, Kaiserslautern, Germany (prof. A.W. Lin).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

6. Efficient Technology for Dealing with Logics and Automata (Not Only) for Formal Analysis and Verification

   Different types of logics and automata are among the most fundamental objects studied and applied in computer science for decades. Nevertheless, there are many unsatisfactorily solved problems in this area, and new, exciting problems related to ever new applications of logics and automata are constantly emerging (e.g., in the formal verification of finite and infinite state systems with various advanced control or data structures, in decision procedures, in program or hardware synthesis, in quantum program verification, or even in methods for efficient search in various types of data or network traffic).

   The subject of the dissertation will primarily be the development of the state of the art in efficient work with various logics (e.g. over pointer structures, strings, various arithmetic, temporal logics, etc.). To this end, approaches based on different types of decision diagrams, automata, but also e.g. approaches based on the existence of a model of bounded size or on efficient reductions between different types of logical theories will be investigated. In connection with this, methods for efficient work with decision diagrams and different types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will be developed. The work will include theoretical research as well as prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be solved in collaboration with the VeriFIT team working on the development of techniques for working with logics and automata and their applications at FIT BUT. There is also the possibility of close cooperation with various foreign partners of VeriFIT: Academia Sinica, Taiwan (Prof. Y. Chen); TU Vienna, Austria (Assoc. F. Zuleger); LSV, ENS Paris-Saclay (Assoc. M. Sighireanu); IRIF, Paris, France (Assoc. A. Bouajjani, Assoc. P. Habermehl, Assoc. C. Enea), Verimag, Grenoble, France (Assoc. R. Iosif); Uppsala University, Sweden (prof. P.A. Abdulla, prof. B. Jonsson); or RPTU, Kaiserslautern, Germany (prof. A.W. Lin).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

7. Formal Methods in Design and Analysis of Quantum Programs

   Development in the field of quantum computers is moving forward unstoppably. However, for the effective use of quantum hardware, it is necessary to have the right support at the software level, i.e., quantum algorithms for solving problems, as well as tools for compiling, optimizing and synthesizing quantum programs, mapping them to a specific quantum technology, mapping them to a fault tolerant protocol, analyzing, simulating, and debugging quantum programs, etc.

   The subject of the dissertation will primarily be the development of current knowledge in the field of formal methods in quantum program design. The main focus here will be on the use of automata theory, automatic reasoning, and decision diagrams. In connection with this, methods for effective work with decision diagrams and various types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will also be developed. The work will include both theoretical research and prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be carried out in collaboration with the VeriFIT team at FIT BUT, which develops techniques for working with logics and automata and their applications. There is also the possibility of close cooperation with various foreign VeriFIT partners: Academia Sinica, Taiwan (prof. Y. Chen); Uppsala University, Sweden (prof. P.A. Abdulla, prof. R. S. Thinniyam); or Leiden Institute for Advanced Computer Science, Leiden, the Netherlands (prof. A. Laarman).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

8. Formal Methods in Design and Analysis of Quantum Programs

   Development in the field of quantum computers is moving forward unstoppably. However, for the effective use of quantum hardware, it is necessary to have the right support at the software level, i.e., quantum algorithms for solving problems, as well as tools for compiling, optimizing and synthesizing quantum programs, mapping them to a specific quantum technology, mapping them to a fault tolerant protocol, analyzing, simulating, and debugging quantum programs, etc.

   The subject of the dissertation will primarily be the development of current knowledge in the field of formal methods in quantum program design. The main focus here will be on the use of automata theory, automatic reasoning, and decision diagrams. In connection with this, methods for effective work with decision diagrams and various types of automata (automata over words, trees, infinite words, automata with counters, registers, etc.) will also be developed. The work will include both theoretical research and prototype implementation of the proposed techniques and their experimental evaluation.

   The work will be carried out in collaboration with the VeriFIT team at FIT BUT, which develops techniques for working with logics and automata and their applications. There is also the possibility of close cooperation with various foreign VeriFIT partners: Academia Sinica, Taiwan (prof. Y. Chen); Uppsala University, Sweden (prof. P.A. Abdulla, prof. R. S. Thinniyam); or Leiden Institute for Advanced Computer Science, Leiden, the Netherlands (prof. A. Laarman).

   Supervisor: Lengál Ondřej, doc. Ing., Ph.D.

9. Human in the loop (HIL) in the adaptation of AI systems

   This PhD topic is driven by the needs of users for which the privacy of data is crucial and concentrates on improving AI (especially speech recognition) system in case the user is able to identify and correct systems errors. The topic includes proper evaluations of HIL systems, selection of data to be proposed for correction and actual fine-tuning / adaptation techniques working with large pre-trained models. 

   Supervisor: Černocký Jan, prof. Dr. Ing.

10. Human in the loop (HIL) in the adaptation of AI systems

    This PhD topic is driven by the needs of users for which the privacy of data is crucial and concentrates on improving AI (especially speech recognition) system in case the user is able to identify and correct systems errors. The topic includes proper evaluations of HIL systems, selection of data to be proposed for correction and actual fine-tuning / adaptation techniques working with large pre-trained models. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

11. Improving human-LLM interaction

    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.

    Supervisor: Móro Róbert, Ing., Ph.D.

12. Improving human-LLM interaction

    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.

    Supervisor: Móro Róbert, Ing., Ph.D.

13. Improving natural language processing

    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. 

    Supervisor: Šimko Marián, doc. Ing., Ph.D.

14. Improving natural language processing

    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. 

    Supervisor: Šimko Marián, doc. Ing., Ph.D.

15. Low-resource text processing with the help of data synthesis

    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.

    Supervisor: Šimko Jakub, doc. Ing., PhD.

16. Low-resource text processing with the help of data synthesis

    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.

    Supervisor: Šimko Jakub, doc. Ing., PhD.

17. Meta-evaluation of multilingual texts by Large Language Models

    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.

    Supervisor: Macko Dominik, doc. Ing., Ph.D.

18. Meta-evaluation of multilingual texts by Large Language Models

    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.

    Supervisor: Macko Dominik, doc. Ing., Ph.D.

19. Multimodal analysis for assessment of mental health

    Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.

    Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.

    Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.

    Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.

    Few Words About Supervision: I have extensive experience of working in the field of neuro-signal and neuroimage processing and I am currently head of a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz

    Supervisor: Malik Aamir Saeed, prof., Ph.D.

20. Multimodal analysis for assessment of mental health

    Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.

    Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.

    Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.

    Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.

    Few Words About Supervision: I have extensive experience of working in the field of neuro-signal and neuroimage processing and I am currently head of a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz

    Supervisor: Malik Aamir Saeed, prof., Ph.D.

21. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

22. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

23. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

24. Performance of voice deepfake detection by humans and machines

    The work will start with getting familiar with the basics of the problem of voice deep fake detection (DFD), terminology, available techniques, data and challenges (AVSpoof, WildSpoof), with the history and state of the art techniques and tools for speaker recognition (wespeaker toolkit), with state of the art techniques and tools for personalized text to speech (pTTS) synthesis and voice conversion. The topic will then concentrate on the evaluation of human and machine performance in DFD, while concentrating on the motivation aspect of human subjects to simulate true attacks on people. The PhD will then progress in both (1) technical (advances in deepfake generation and detection technology) and (2) human-aspects. The work counts on collaboration with specialists from psychology and sociology. 

    Supervisor: Černocký Jan, prof. Dr. Ing.

25. Recommender and adaptive web-based systems

    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.

    Supervisor: Kompan Michal, doc. Ing., PhD.

26. Recommender and adaptive web-based systems

    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.

    Supervisor: Kompan Michal, doc. Ing., PhD.

27. Research and advancement of the visualization, interaction and AI technique Malleable Glyph

    The goal of the doctoral program will be to develop the newly proposed technique, Malleable Glyphs, in the fields of human-computer interaction, visualization, and computer vision. The researcher will develop machine learning and computer vision algorithms for glyphs. The researcher will also co-organize The Malleable Glyph Challenge, design new glyph modalities, evaluate glyphs submitted by students and individuals from third-party institutions, and create a glyph taxonomy.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

28. Research and advancement of the visualization, interaction and AI technique Malleable Glyph

    The goal of the doctoral program will be to develop the newly proposed technique, Malleable Glyphs, in the fields of human-computer interaction, visualization, and computer vision. The researcher will develop machine learning and computer vision algorithms for glyphs. The researcher will also co-organize The Malleable Glyph Challenge, design new glyph modalities, evaluate glyphs submitted by students and individuals from third-party institutions, and create a glyph taxonomy.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

29. Speech recognition and data processing for ATC-pilot communication in aviation

    This PhD topic is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology), language model as well as interaction with sources of meta-data (radar information, airport). Special attention will be paid to the use of large pre-trained model, code-switching, reliable language identification from short segments and adaptation to the conditions of given airport with crawled data.

    Supervisor: Černocký Jan, prof. Dr. Ing.

30. Speech recognition and data processing for ATC-pilot communication in aviation

    This PhD topic is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology), language model as well as interaction with sources of meta-data (radar information, airport). Special attention will be paid to the use of large pre-trained model, code-switching, reliable language identification from short segments and adaptation to the conditions of given airport with crawled data.

    Supervisor: Černocký Jan, prof. Dr. Ing.

31. User Activity Prediction in Social Media Environments

    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.
    • 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 SIGIR '25. ACM, New York, NY, USA, 3357–3366.

    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.

    Supervisor: Srba Ivan, Ing., Ph.D.

32. User Activity Prediction in Social Media Environments

    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.
    • 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 SIGIR '25. ACM, New York, NY, USA, 3357–3366.

    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.

    Supervisor: Srba Ivan, Ing., Ph.D.

33. User Experience in Modern User Interfaces

    The goal of the doctoral study is to analyze the characteristics of modern and popular user interfaces, with a focus on cognitive load, addiction-inducing techniques, the perception of AI-generated content, the perception of advertising, and other significant contemporary phenomena. The student will prototype user interface elements and design and conduct user tests that will enable a deeper understanding of the phenomena under study, highlight existing risks, and lead to the design of user interface elements that will represent alternatives beneficial to the user.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

34. User Experience in Modern User Interfaces

    The goal of the doctoral study is to analyze the characteristics of modern and popular user interfaces, with a focus on cognitive load, addiction-inducing techniques, the perception of AI-generated content, the perception of advertising, and other significant contemporary phenomena. The student will prototype user interface elements and design and conduct user tests that will enable a deeper understanding of the phenomena under study, highlight existing risks, and lead to the design of user interface elements that will represent alternatives beneficial to the user.

    Supervisor: Herout Adam, prof. Ing., Ph.D.

35. Using machine learning to solve energy and environmental problems

    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

    Supervisor: Rozinajová Věra, Doc., Ph.D.

36. Using machine learning to solve energy and environmental problems

    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

    Supervisor: Rozinajová Věra, Doc., Ph.D.