Bachelor's · Master's · PhD at FIT CTU

Thesis and PhD topics

Topics from my research lines: recommender systems and search, safe language models and AI agents, AI for climate and science, and AI for learning. Every thesis, the bachelor's ones included, should have research potential and a result that can be measured.

The topics below are examples. Your own topic is just as welcome, especially one that builds on this research and takes it further. Propose a topic

Browse topics How to start

Updated October 2026 · Supervisor: doc. Ing. Pavel Kordík, Ph.D., Department of Applied Mathematics, FIT CTU in Prague

From a thesis to a paper

A few students who started with a thesis in the group.

Čeněk Žid
Master's thesis on recommending opportunities to students (2022) paper at CISIS 2023 PhD, paper at CIKM 2025
Vojtěch Cahlík
Master's thesis on neural networks that resize themselves (2022) paper at IEEE CSIT 2022 PhD on explaining language models (xAI 2025)
Vojtěch Vančura
PhD on scalable recommender models (2025) papers at RecSys 2022 and 2024 and in ACM TORS; the ELSA and beeFormer code is open-sourced by Recombee
Tomáš Řehořek
PhD on recommendation models (2019) co-founder of Recombee, which he now leads

Theses and papers since 2002

Each brick is a defended thesis (light) or a paper (solid), coloured by topic. Hover a brick, or click a year to see what students worked on.

Topics

Most topics can be scaled to a bachelor's or master's thesis, and some to a PhD. The level badges are a suggestion, not a rule.

Recommender systems & search

With the Recombee research team, on production-scale data.

MScPhD

Recommender models that read

Clicks tell you what people like, text tells you what things are. Combine both so that a brand-new item gets good recommendations from day one.

Train shallow autoencoders and sentence encoders on interaction data; test them on public benchmarks and on Recombee-scale catalogues.

Measured by: recall and NDCG on cold-start items against ELSA and beeFormer baselines.

Builds on: ELSA (RecSys 2022), beeFormer (RecSys 2024), Language Embeddings Meet Shallow Autoencoders (UMAP 2026)

beeFormer ELSAI'm interested
BScMSc

Sparse embeddings and fast retrieval

Dense vectors are accurate but expensive. How sparse can an embedding get before recommendations and search get worse?

Compress embeddings, build an index and map the trade-off between memory, speed and quality.

Measured by: quality versus memory and latency curves on public datasets.

Builds on: The Future is Sparse (RecSys 2025), SparseELSA (ACM TORS 2025)

SparseELSAI'm interested
MScPhD

Simulated users for testing recommenders

Offline metrics often disagree with A/B tests. Can language-model agents that behave like users close that gap?

Build user simulators, calibrate them on logged behaviour and check whether they rank models the way real traffic does.

Measured by: rank correlation between simulated and online results.

Builds on: best paper at EvalRS @ KDD 2023 on offline–online evaluation

I'm interested
MScPhD

Recommendation for AI agents

More and more shopping and searching is done by agents on behalf of people. What should a recommender show to an agent, and how do we keep it honest?

A PhD line that started in 2026; a master's thesis can take one piece of it, such as an agent-facing recommendation API or a benchmark.

Measured by: welfare of both sides, communication cost and robustness to manipulation in simulated markets.

Builds on: Evaluating the Division of Labour in Strategic Two-Sided Agentic Marketplaces (IJCCI 2026)

I'm interested

Safe and useful language models

Guardrails, explanations and safety data for systems people use.

BScMSc

Guardrails for children's AI tutors

Children are not small adults, and general-purpose moderation misses what matters for them. Make the Tiny chatbot used in Czech schools safer.

Extend a child-centred safety taxonomy, generate test conversations and train small guard models that run cheaply.

Measured by: recall and precision on the child-safety benchmark, and cost per conversation.

Builds on: Mind the Alignment Gap (WOAH @ EMNLP 2026)

Tiny talkI'm interested
MScPhD

Explanations you can check

A language model can explain its answer convincingly and still be wrong about why it answered. Make explanations grounded in the model's actual reasoning.

Design tasks where the right reasoning is known, fine-tune models and test how faithful their explanations are.

Measured by: faithfulness of explanations on tasks with known reasoning.

Builds on: Reasoning-Grounded Natural Language Explanations for Language Models (xAI 2025)

I'm interested
BScMSc

Safe semantic search

Semantic search finds what you mean, including things a marketplace should not show. Detect harmful and sensitive queries without blocking normal ones.

Extend the SHIELD dataset to new languages or domains and compare fast retrieval-based classifiers with fine-tuned models.

Measured by: detection quality against false blocks of normal queries.

Builds on: Mitigating Risks in Marketplace Semantic Search (UMAP 2025)

SHIELDI'm interested

AI for climate, regions and science

In the EU projects FOCAL and PoliRuralPlus and with partners such as Meteopress and HZDR.

MScPhD

Sharper climate projections and nowcasting

Climate models work on coarse grids, but a town needs to know about its own valley. Downscale projections to 1 km, or predict rain from radar for the next hour.

Train super-resolution or diffusion models on climate and radar data and validate them against observations.

Measured by: error against station and radar observations, with calibrated uncertainty.

Builds on: NeurIPS 2025 climate workshop paper; 1st place in Weather4cast 2022

FOCALI'm interested
BScMSc

A geo-aware AI assistant for rural planning

Advice that is right in Catalonia can be harmful in Bohemia. Make an assistant that grounds its answers in local maps and data, and knows when to refuse.

Work with the JackDaw agent and the geo-safety dataset: retrieval over spatial data, evaluation on real questions from the regions.

Measured by: share of harmful advice caught and of correct, grounded answers.

Builds on: No Safe Words (GeoAI 2026), JackDaw (FedCSIS 2025)

PoliRuralPlusI'm interested
MSc

Recommending models to scientists

Picking the right closure model for a fluid simulation is like picking a film: there are many, and past runs tell you which fit. Treat it as a recommendation problem.

Apply matrix completion and cold-start recommenders to simulation results, together with researchers from HZDR in Dresden.

Measured by: how often the recommended model beats the standard choice.

Builds on: AI Thermal Fluids 2025; arXiv 2026

arXivI'm interested

AI for learning

Tools that already have readers and classrooms, studied with real users.

BScMSc

Interactive slides in school lessons

Lessons for the Tiny classroom assistant are built in SlAIdy and used in Czech schools. What makes interactive slides work in a primary or secondary school classroom, and how should the interface change for teachers and pupils?

Adapt the user experience for teachers (preparing, presenting, classroom interaction) and run a user study in schools.

Measured by: preparation time, pupil engagement or quiz results, compared with ordinary slides.

Builds on: SlAIdy; Tiny, with 200+ registered schools

SlAIdy TinyI'm interested
BScMSc

A book that recommends itself

The p-book teaches recommender systems and is personalised by one. Does reading by missions, feeds or a map change what people learn?

Design and run a controlled study in the live book, or generate new versions of a chapter for a reader and test whether they are correct and helpful.

Measured by: what readers learn and how long they stay, by reading mode.

Builds on: p-book demo at RecSys 2026

p-bookI'm interested
BScMSc

How reliable are AI-drafted research gaps?

The RecSys Atlas puts 2,200 papers on one map, with open problems drafted by AI from the abstracts. Which of those problems are real?

Check AI-drafted claims against the papers and against expert judgement, then improve the method or apply it to another field.

Measured by: precision of the drafted gaps against expert ratings.

Builds on: RecSys Atlas (2026), 20 years of ACM RecSys

AtlasI'm interested
BScMSc

Matching students with real projects

Companies post projects, students have skills and timetables. A match only works if both sides say yes, at the right moment.

Work on reciprocal recommendation and outreach timing for the EduMatch platform.

Measured by: response and acceptance rates in offline replay and in an online test.

Builds on: CISIS 2023, Active Recommendation for Email Outreach Dynamics (CIKM 2025)

EduMatchI'm interested

Your own topic

A topic of your own is welcome if it builds on the research above and takes it further. Write what you want to find out and how you would measure it.

Propose a topic

How to start

1

Pick a topic

Or two, or bring your own that builds on the research above. You don't need to know everything about it yet.

2

Send a short message

Use “I'm interested” on a topic: the form opens with the topic filled in. Name, e-mail and a few lines are enough.

3

Meet for half an hour

We shape the assignment together, agree on who from the team helps you, and register the thesis in KOS.

When should I write?

Ideally one semester before you want to register the thesis, so there is time to try the data and tools. Later works too.

Do I have to study at FIT CTU?

Bachelor's and master's theses are mostly for FIT students; students of other faculties can agree on external supervision. For a PhD, see the doctoral study programme at FIT CTU.

How do I apply for a PhD?

Write through the form on my homepage: which research line interests you, what you have done so far (company, projects, papers) and a link to your CV or LinkedIn. We meet, agree on a topic, and you then apply to the doctoral programme at FIT CTU.

Which language?

We work in English: papers, code, data and meetings with partners. The thesis itself can be in Czech or English; in English it is easier to turn into a paper.

Who will I work with?

With me and usually with one of my PhD students or postdocs who knows the topic in detail; on Recombee topics also with its research team.

Not sure which topic? Write anyway.

Tell me what you enjoy and what you would like to find out, and we will look for a topic together.

Get in touch