Open Science: From Principles to Daily Research Practice

This session briefly introduces key elements of open science in everyday research practice, including open licences, trusted repositories, and the main differences between institutional and disciplinary repositories. It also addresses research data management, with a focus on data management plans and practical support through tools such as the Data Stewardship Wizard. The session concludes by outlining the types of open science support available at university and faculty level and where researchers can seek practical guidance.

The slides from the talk are available at https://is.muni.cz/auth/el/fi/jaro2026/IA067/um/slides/ (after logging in to IS).

Lab Fest, held as part of the Computer Science Colloquium, aims to familiarise the academic community with the activities of research groups at the faculty, specifically their staffing, areas of interest, involvement in grants, and ongoing collaborations at the university, within the Czech Republic and internationally; their future direction and current research and development outcomes.

AGDM: Studio of Graphic Design and Multimedia

EMLAB: The Design and Architecture of Digital Systems Laboratory

How Algorithm Design Can Save Participatory Budgeting

In a participatory budgeting process, the aim is to select which projects, from those submitted by citizens, should be implemented in a given city. The city sets a budget, people submit project ideas—such as building cycle paths, organising various festivals, or offering language classes—each with a different cost, and eventually the citizens vote on which of these projects are carried out. Whilst many cities implement participatory budgeting, most of them use very basic voting rules that do not ensure proportional representation of the voters and that have a number of drawbacks. For example, a relatively small, coordinated group of citizens can hijack the entire budget. In this talk, I will present the Method of Equal Shares—a voting rule recently proposed by Pierczyński, Peters and Skowron—which offers strong guarantees of proportionality, and I will recount the story of how this rule evolved from a theoretical concept to being applied in practice. This story involves a number of algorithmic and computational complexity developments that were necessary to make it feasible (in particular, to explain the sometimes surprising results that the rule can produce). During the talk, I will also discuss several other participatory budgeting rules and compare them with the Method of Equal Shares.

Challenges in Data Visualisation and Building Trust in It

Data visualisation is a very powerful tool for analysing and understanding complex data. However, there are many factors that play a crucial role in this process. These include not only the size and dimensionality of the data, but also the design decisions underpinning appropriate visual representations and interactions with them, the limitations of human cognitive perception, and the ability of diverse target audiences to interpret the representations correctly. All these aspects significantly influence the stakeholders’ trust in the visualisation. In this talk, I will present the main aspects and challenges associated with data visualisation, linked to the three main pillars: data, users, and visuals that enable insight. The talk will conclude with several observations and recommendations for interdisciplinary collaboration aimed at building mutual trust between project partners.

This talk forms part of the application process for a professorship.

Learning to Work with AI: Lessons from a US Tech Start-up

The way knowledge workers operate has fundamentally changed over the last three years. In software engineering, AI coding tools have evolved from a novelty to an everyday part of the infrastructure – and AI skills are increasingly expected and assessed in job interviews. But what does this actually look like in practice, and what should it mean for the way we teach computer science?

Drawing on the experience of adopting AI tools at a US fintech start-up over the past two years, I’ll share some lessons my company learnt the hard way: why adoption is harder than it looks, how documentation debt becomes your AI’s blind spot, and why some languages and frameworks lend themselves to AI-assisted development far better than others. Nobody knows what’s coming next, but I’ll share my perspective on what it takes to be an effective software engineer in this new world.

Principles of Accessible Teaching

We will focus on universal design and how to ensure that teaching is accessible to all students, paying particular attention to hidden disabilities (such as psychological issues and illnesses, the number of which has risen dramatically over the past five years) at different stages of the semester, with a particular focus on the examination period (which is fast approaching). During the discussion, we will present the universal design framework currently being developed at the Teiresiás Centre, which will serve as a practical guide for all teachers, regardless of their years of teaching experience.

CoFI break

Neural Networks as Parameterised Gradient-based Solvers

Neural networks were long regarded primarily as pattern recognition systems that are effective at detecting statistical regularities in data, but ill-suited to problems requiring reasoning. When they did appear to solve such tasks, a common explanation was that they were simply memorising the training set. The recent success of large language models has challenged this view, as these models can solve difficult and novel mathematical problems. Yet we still understand surprisingly little about how they do so.

In this talk, I present a mechanistic perspective on neural reasoning through the lens of graph neural networks (GNNs). Specifically, I show how GNNs trained for Boolean satisfiability (SAT) can be interpreted as parameterised primal–dual algorithms that optimise a Lagrangian relaxation of a SAT instance. This viewpoint links learned message-passing dynamics to classical optimisation and provides a concrete framework for analysing when and why these models find satisfying assignments in constraint satisfaction problems. I conclude by discussing what this perspective suggests for the design of more efficient language models.

Verse – designing a new programming language for the Metaverse

Verse is a new programming language being developed at Epic Games for building interactive experiences in the metaverse. Unlike traditional scripting languages for game engines, Verse is designed as a functional logic programming language with a rigorous theoretical foundation—the Verse calculus (VC), a core calculus for deterministic functional logic programming published at ICFP 2023. In this talk, I will discuss the design philosophy behind Verse and the key challenges of creating a language for a future where code runs in a single global simulation. Verse combines ideas from functional programming (immutability, higher-order functions, strong typing) and logic programming (unification, choice, backtracking), whilst introducing novel features such as failable expressions, speculative execution with transactional rollback, and language-level concurrency. The language aims to be “metaverse first”, with strong compatibility guarantees and an effect system that tracks side effects to ensure safe concurrent execution. I will explain how these theoretical concepts translate into practical language features and discuss the trade-offs we faced in designing a language that must be both accessible to game developers and capable of expressing complex concurrent and reactive systems. The talk will also touch on our experience deploying Verse in Fortnite’s Unreal Editor and our vision for making Verse an open standard for metaverse development.

Towards Verification via Transformation

Following a brief overview of the history of software model checking, we conclude that over the past 25 years we have moved from a lack of tools to an abundance of them. Our vision is to use existing verification tools as components or libraries, just as we now use SMT solvers. New (meta) verification tools should form a new layer on top of the existing architectural layers. This requires clear interfaces and separation of concerns, as well as exchange formats for verification artefacts such as intermediate or partial results. Only by combining verification tools and their underlying technology can we harness the full potential at our disposal. We outline a few existing transformation approaches to demonstrate that achieving this goal is possible in the near future.

Cancelled, postponed until autumn 2026

Jan Jedelský: Transductions between sparse graph classes

The first-order (FO) model-checking problem asks, given a graph G and an FO formula 𝜙, whether G is a model of 𝜙. Using FO logic, one can formulate the problems of determining whether G contains a clique of size k, a distance-d independent set of size k, or a subgraph isomorphic to a graph H; and even whether the tree-depth of G is at most k.

An FO transduction is a notion of non-deterministic reduction for this problem. Informally speaking, a transduction from a graph class 𝒞 to a graph class 𝒟 demonstrates that every graph G∈𝒟 can be obtained from some colouring of some graph H∈𝒞 using FO logic. Hence, solving the FO model-checking problem restricted to inputs from 𝒞 is no easier than solving the same problem restricted to inputs from 𝒟.

In this talk, we characterise transductions between sparse graph classes in purely combinatorial terms. As a corollary of our characterisation, for every k ∈ ℕ, we prove that the class of all graphs with tree-width ≤ k + 1 is not transducible from the class of all graphs with tree-width ≤ k.

Anna Řechtáčková: Code Quality Defects and How to Find Them Automatically

The ability to write code is not the same as the ability to write it well. I study code quality defects found in code written by novice programmers and explore ways to detect them automatically — whether through static analysis or large language models.

In this talk, I will first define code quality defects and present our catalogue of over 100 specific defects. I will then share results from a survey of lecturers at the Faculty of Informatics (FI) and abroad on how they perceive the severity of these defects. Next, I will introduce EduLint, an educational linter I have developed, and describe its integration into the CS1 course IB111 taught at the Faculty of Informatics, Masaryk University (FI MU). I will outline our research into detecting poor identifiers using large language models and compare their performance with that of human evaluators. Finally, I will describe our efforts to detect duplicate code using both static analysis and LLM-based methods.