Back to the main page of "Praxis der Forschung"

Summer semester 2026

Projektgruppe "Praxis der Forschung" / Project Group "Research Practice"

Announcements:

The modules can now be done both in German and English and in both M.Sc. programmes (Informatik (de) and Computer Science (en))
22 April 2026: Information event (details below)
29 April 2026: Kickoff event, first meeting (see below)
Calendar for the first semester
Further participation information for the events in the ILIAS course, please check regularly!

Offered topic areas in Summer semester 2026

The participating research groups for Summer semester 2026 are:

The following list gives an overview of topics that are offered this semester. If you are interested in a topic or have questions, you can contact the relevant staff member. This list will be supplemented before the semester begins. The final list of offered topics will be announced at the information event and the topic presentation at the beginning of the semester (see above) and can afterwards be read on the published slides.

Note: Based on experience, it is worth getting in touch with the respective staff members in advance to avoid disappointments during the topic presentation. If you are interested in carrying out a project within Practice of Research, but your desired topic is not listed here, topics can in exceptional cases still be registered until the end of the first lecture week. Please contact the relevant staff members or research groups (whose topics interest you) early to agree on a topic, and ask them to contact us as soon as possible if they are interested.

List of projects for Summer semester 2026

#	Project	Research group(s)	Supervising staff
1.	Automatisierung von Design-by-Contract bei reaktiven SystemenZiel. Hier wollen wir einsteigen, indem wir (a) ein Verfahren aufstellen, mit welchem Minen und Lernen von Spezifikation gelingt, und (b) eine Hilfsspezifikation für die Systemgrenzen automatisch inferieren. Als Spezifikationssprache soll auf Vertragsautomaten aufgebaut werden, und dadurch ebenso auf das Automatenlernen (vgl. L-Algorithmus von Alduin). Am Ende könnten ein Ansatz und ein Werkzeug entstehen, mit welchem ein Systementwurf vollautomatischer auf funktionale Korrektheit geprüft werden kann. Bemerkung:* Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	KASTEL Beckert	Alexander Weigl Andreas Bremer
3.	Combining rule-based consistency preservation and contract-based consistency repairThis project aims to integrate Consistency Preservation Rules (CPRs) with contract‑based consistency repair by leveraging sophisticated planning tools on graph‑based model representations, thereby enabling scalable, automated management of inter‑view consistency in cyber‑physical system (CPS) development. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	KASTEL Beckert	Mattias Ulbrich
4.	Participatory Design of Digital Tools to Support Children's Disengagement From GamesThis project will explore the participatory design of a digital tool to support children's disengagement from games, which are currently only designed with parental needs in mind. The project will comprise the following steps: (1) Review of existing solutions to support children's disengagement from games, e.g., parental timers, as well as review of literature on parental moderation of children's media use. (2) Participatory design workshops involving children aged 8 to 12 (N=16) and parents (N=16) to understand their perspectives on currently available tools, and ideation of future solutions that protect children's player experiences and autonomy while aligning with parental limits for playing time. This involves the development of (experience) prototypes of digital support tools to support for critical appraisal as part of the workshops. (3) Construction of a design framework for digital tools to support children's disengagement from games based on the outcomes of (1) and (2). Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAR Gerling	Zeynep Yildiz Kathrin Gerling Meshaiel Alsheail
6.	Advanced Strategies for Malleable SAT SolvingSAT solving is highly relevant for a wide range of applications (verification, security, XAI, …). Distributed SAT solving, e.g., on supercomputers or in clouds, achieves best efficiency with scheduling strategies that exploit malleability, i.e., a parallel task's ability to handle fluctuating computational resources during its execution. So far, malleable SAT solving works by (a) initializing new solvers when a task expands onto new resources and (b) suspending/terminating solvers when a task shrinks. This performs well if a task expands monotonically and/or if the input is unsatisfiable. However, satisfiable inputs hardly profit from fluctuating solvers. We aim to investigate advanced methods for malleable SAT solving, such as welcoming joining solvers with helpful data, transferring a leaving solver’s progress to the remaining solvers, or serializing SAT solver states to disk or over the network to reuse them later. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	ITI Schreiber	Dominik Schreiber
7.	Continual Learning in LLM-powered AgentsDiese Arbeit beleuchtet das kontinuierliche Lernen von Agenten, die auf Large Language Models basieren. Im operativen Einsatz erhöhen diese Agenten ihre Leistungsfähigkeit, indem sie über veraltete Wissensstände und Fähigkeiten hinaus aktuelle, relevante Informationen einbinden, die ursprünglich nicht im Trainingsdatensatz enthalten waren. Dabei spielen die Techniken Retrieval-Augmented Generation (RAG) und Test-Time-Training (TTT) eine entscheidende Rolle, denn sie ermöglichen es, neue Daten in den Entscheidungsprozess einzubeziehen. RAG nutzt externe Datenquellen, ohne dabei die Modellparameter anzupassen und stellt sie im Kontextfenster zur Verfügung, wohingegen TTT eine dynamische Anpassung der Parameter in Echtzeit erlaubt. Ziel dieser Arbeit ist es, zu untersuchen, wie beide Methoden effektiv kombiniert werden können und wann welche Lerntechnik eingesetzt werden sollte, um die Erfolgsquote auf verschiedenen Aufgaben zu erhöhen. Die Lerntechniken sollen so kombiniert werden, dass die jeweiligen Vorteile maximal genutzt werden. Um das genannte Ziel zu erreichen, werden die jeweiligen Lerntechniken einzeln analysiert, um entsprechende Vor- und Nachteile herauszuarbeiten. Hierfür werden bestehende Datensätze als Grundlage für die Untersuchung genutzt. Zusätzlich wird ein eigener Datensatz erstellt, welcher Fokus auf die besonderen Herausforderungen des kontinuierlichen Lernens legt. Als letzten Schritt der Arbeit soll ein hybrider Ansatz entwickelt werden, der beide Lerntechniken in Bezug auf Leistung und Effizienz kombiniert. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAR Niehues	Lukas Hilgert
8.	Graph Neural Network Limitations in Power Systems: From Spectral Analysis to ExplainabilityPower systems demand scalable computational methods for critical tasks such as cascading failure prediction, dynamic stability assessment, AC power flow calculation, and wind power forecasting, for which Graph Neural Networks (GNNs) are a natural fit since the power grid is inherently a graph. However, the message-passing mechanism introduces fundamental limitations: oversmoothing, where node representations become indistinguishable with depth; oversquashing, where information through sparse connections gets distorted; and limited expressiveness, where structurally different configurations appear identical to the model. Recent work has shown these limitations are more nuanced than previously assumed and depend critically on graph topology and task structure. Energy graphs represent an underexplored and structurally diverse graph class, ranging from physically grounded network representations to more abstract relational structures depending on abstraction level and modeling choices. In this work, we aim to provide the first systematic quantification of GNN limitations across power system tasks spanning node, edge, and graph prediction levels, using spectral properties of energy graphs to derive analytical predictions verified empirically through explainability methods (XAI). We compare findings against traffic networks, molecular graphs, and social networks, and assess architectural remedies tailored to each limitation, including Graph Transformers for long-range dependencies, higher-order GNNs for expressiveness, and graph rewiring or normalization for oversmoothing. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAI Schäfer	Martin Sadric
9.	Reconfiguration of GraphsGiven a set of combinatorial objects U and a set of operations F on these objects, the reconfiguration graph R_F(U) is defined as the graph with vertex-set U, i.e., the combinatorial objects in U are vertices of R_F(U), and an edge between two vertices U_1 and U_2 if and only if there is an operation f in F such that f(U_1)=U_2. The properties of reconfiguration graphs have been extensively studied for various combinatorial objects, such as triangulations in the plane, spanning trees of fixed point sets in the plane, as well as drawings, independent sets, or colorings of a fixed graph, just to name a few. However, little is known for the case that U is a class of graphs. We want to study the operations F of adding or removing a single edge from a graph in U, and investigate properties of the resulting reconfiguration graph. That is, for different graph classes G, we want to examine the graph R_F(U) with vertex-set G, and an edge between two vertices G_1 and G_2 if and only if G_2 can be obtained by adding an edge to G_1 or deleting an edge from G_1. Possible candidates for the graph class G include chordal graphs, perfect graphs, interval graphs, comparability graphs, as well as graphs of treewidth k or treedepth k for some fixed integer k. We are in particular interested in properties of the resulting reconfiguration graphs that are frequently considered in this field. This includes the connectivity, i.e., whether the reconfiguration graph is connected and if not how many components it has, the diameter and radius, as well as the computational complexity to compute a shortest path between two given vertices. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	ITI Ueckerdt	Torsten Ueckerdt Samuel Schneider
10.	Dependable and Trustworthy Embedded Systems for Space ApplicationsEmbedded systems have reached a ubiquitous position. They perform tasks not just in consumer electronics but also in critical fields such as medical devices and avionics. Malfunctions can cause serious damages, injury or death. Especially space applications become an increasingly relevant topic for European and German ambitions. Compared to other fields, space applications require utmost dependability and trustworthy. Therefore, the systems must cope with different failure and attack scenarios. This project's goal is to prototype dependable and trustworthy embedded systems for space applications. One area of interest is to compare different operating systems or FPGA implementations to realize the communication between redundant modules. Interesting operating systems could be Linux and Zephyr. Linux would be an especially interesting contestant as Linux' real-time scheduling (PREEMPT_RT) is a new addition. Zephyr is a relatively recent development and sees rapid growth. Therefore, it would be interesting to see if it can be a part of critical space applications. Another interesting area is the automatic verification of hardware designs and system health. Such an approach could speed up the development of more dependable embedded systems. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	ITEC Henkel	Hassan Nassar
11.	Combinatorial Pattern Matching for Multidimensional DataThe task for this PdF project is to investigate the complexity of combinatorial pattern detection problems for multidimensional data, e.g. detecting four 1s forming a square in a Boolean matrix. For which formulations can we improve over exhaustive search algorithms, for which can we prove conditional lower bounds? Specific formulations include other patterns than squares, generalizations to higher dimensions than two-dimensional matrices, further degrees of freedom (such as rotations of patterns), as well as related settings (e.g., counting occurrences of patterns, etc.). Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	ITI Künnemann	Geri Gokaj Marvin Künnemann
12.	Effiziente Darstellung von kolloidaler DispersionKolloidale Dispersion ist ein Phenomaen, was ungewoehnliche Beleuchtungseffekte erzeugen kann. Ein prominentes Beispiel ist ein Kelch, welcher je nach Belichtungsseite die Farbe aendert. Dies ist nicht durch die herkoemmliche Renderinggleichung und die Materialmodelle darin abgedeckt. Ziel dieses Projektes ist, die chemischen und physikalischen Zusammenhaenge hier fuer die Grafik aufzubereiten, und ein Materialmodell und Erweiterung der Renderinggleichung zu erforschen um solche Effekte nachbilden zu koennen. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IVD Dachsbacher	Johannes Schudeiske
13.	Detecting and Localizing Hardware Faults in Neural Networks and Language ModelsThis project focuses on localizing hardware faults in neural networks and small language models running on AI accelerators. Such faults can flip stored weight bits, corrupt activations, or make processing elements stuck, often causing silently wrong outputs instead of a crash. The project first generates input test patterns to detect faulty behavior, known as ATPG (Automatic Test Pattern Generation), and then focuses on: localizing the fault to a model component such as a layer, neuron, weight group, or attention head, related to DTPG (Diagnostic Test Pattern Generation). The work will be done in software by simulating hardware faults in PyTorch, starting with CNN-based models and then moving to small transformer models. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	ITEC Tahoori	Tara Gheshlaghi
14.	Reinforcement Learning for Vision-Language-Action ModelsVision-language-action models (VLAs) transfer common world knowledge from vision language models to the domain of robotic control through large-scale pre-training, promising better generalization capabilities than task-specific policies trained from scratch. While such models produce impressive-looking demos of complex tasks such as cloth folding, they rely on substantial amounts of task-specific demonstration data for additional finetuning. Furthermore, VLAs still struggle with high-precision tasks, making them unreliable for large-scale deployment in the real world. VLAs inherently rely on imitation learning, where humans demonstrate a task, typically by teleoperating a robot, and the policy is trained to imitate the demonstrated behavior using standard supervised learning. In contrast, reinforcement learning (RL) allows the robot to explore the environment on its own and find an optimal solution to the task at hand. RL does not need expert demonstrations, but only scalar rewards. In this project, we want to investigate to what extent RL can a) reduce or eliminate the need for task-specific demonstrations and b) improve task success rate in cases where teleoperated demonstrations alone are insufficient. Examples include high-precision tasks such as threading a needle and tasks involving dexterous hands, which are unintuitive for humans to teleoperate. A major challenge of RL is its data inefficiency: Typically, many thousand or even million environment interactions are required to discover good actions. To this end, we will utilize the supervised pretraining of VLAs, which equips them with general world knowledge and, ideally, good behavioral priors for many different tasks. We will investigate various approaches to VLA RL finetuning, including residual policies, noise-space policies, and full-model finetuning, evaluate their tradeoffs, and then focus on one of them for the remainder of the project. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAR Neumann	Florian Seligmann Emiliyan Gospodinov
15.	Membership Inference Attacks on Gait RecognitionMembership inference attacks (MIA) allow an attacker to infer whether a specific record was part of a machine‑learning model's training set, thereby exposing personal information such as health status or location history. This type of privacy breach has been demonstrated on various machine learning tasks, including face and voice recognition, but is still missing for gait recognition. At the same time, gait has been shown to enable numerous sensitive inferences, exacerbating the privacy risk of leaked training data. In this Practice of Research topic, students will first survey state‑of‑the‑art MIA methods used for other biometrics, then methodically design novel MIAs for deep learning gait recognition models. After implementing the attack(s), they should be rigorously evaluated to investigate to which extent MIA on gait recognition is possible and which factors influence their success. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	TM Strufe	Julian Todt
16.	Kinesthetic Teaching and Error Detection on Humanoid Robotst.b.d. Bemerkung: IAR Mombaur Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAR Mombaur	Björn Plonka

Topics that remained open

Not all offered topics were taken. If you are interested in a topic that remained open, you can contact the supervising person(s) if you would like to take it over in the next semester (if it is offered again.)

#	Project	Research group(s)	Supervising staff
2.	Lenses for Models, CategoricallyModels in Model-driven engineering are abstractions used to manage the complexity of system development: each model describes a different purpose-oriented aspect of the system. However, such models are never fully separate: dependencies and overlap abound between them. Lenses are an idea from the field of software engineering and functional programming: a lens consists of a get : M → V and a put : V × M → M operation which together satisfy the “lens laws”: put(get(m), m) = m get(put(v, m)) = v Two research questions could serve as a starting point for a PdF/thesis project: 1. Lens construction: Lenses on trees are usually defined using combinators. Can similar combinators be defined for models (formalised as attributed typed graphs)? 2. Functionalisation: the Reactions language of the Vitruv tool is a Java-based DSL for describing model updates in the Vitruvius tool. Could its programs be “functionalised” via a translation to such lens combinators?	KASTEL Beckert	Terru Stübinger
5.	Geometric Gaussian Mixture Learning of ManifoldsIn many applications, it is useful to represent a deterministic manifold—a structured, lower-dimensional surface in a higher-dimensional space—in a probabilistic way. Although the manifold itself is exact, a probabilistic representation provides robustness to noise and enables the use of statistical inference tools. A common approach is to approximate the manifold using a Gaussian mixture model (GMM), which expresses a complex distribution as a weighted sum of Gaussian components. This allows flexible modeling of nonlinear and multimodal structures by concentrating probability mass around the manifold. Gaussian mixtures are typically fitted using the Expectation-Maximization (EM) algorithm. EM alternates between assigning data points to components (expectation step) and updating the parameters of those components (maximization step). While widely used, EM has several limitations. It is not guaranteed to find the global optimum and is highly sensitive to initialization, often converging to suboptimal local solutions. Additionally, GMMs are data hungry, especially in high dimensions where estimating covariances reliably becomes difficult. The algorithm is also computationally expensive, as each iteration scales with both the number of data points and mixture components. Now enters the Geometric learning of the Gaussian Mixture of Manifolds.	IAR Beyerer	Ali Darijani

Events and dates

The methods events will probably take place according to the following schedule. Attendance at the first event is, among other things, a prerequisite for participation in Praxis der Forschung; important organisational information will be announced there.

Information event

On 22 April 2026 the information event for Praxis der Forschung / Research Practice will take place:

At 12:30 in the informatics library (ground floor bldg 50.34) the organisational basics regarding the procedure and design of PdF will be discussed.
Between 13:00 and 14:00 there will be a poster show of the semester's PdF topics at the same place in the infromatics library (bldg. 50.34)

As indicated in the calendar section below, the first event will already be on 29 April: A short kickoff sessions and a coordinating event for the presentation seminar.

First semester

Date	Time	Mode	Event	Location	Lecturer	Notes
2026-04-22	12:30 – 14:00		Information Event	Informatik-Bibliothek (bldg 50.34)	M. Ulbrich
2026-04-29	15:00 – 16:00	All	Presentation Workshop Preparation	online (Zoom)	Frau Jüttner	Assigning groups
2026-04-30	10:30 – 11:00	All	Kick off	SR Forum (bldg 30.95)	M. Ulbrich
2026-04-30	11:00 – 16:00	All	Project Management	SR Forum (bldg 30.95)	Dr. Lehr
2026-05-12	14:00 – 15:30	All	Literature Research & Citations	Seminarraum 34, 3. OG Altbau (Bibliothek Süd).	Fr. Sielaff	Changed date!
2026-05-13	15:00 – 18:30	Groups	Presentation Workshop I	SR 006/104 (bldg 30.96)	Frau Jüttner
2026-05-19	14:00 – 16:00	Coaching	Project Management	online	Dr. Lehr
2026-05-20	15:00 – 18:30	Groups	Presentation Workshop I	SR 006/104 (bldg 30.96)	Frau Jüttner
2026-05-21	14:00 – 16:00	Coaching	Project Management	online	Dr. Lehr
2026-06-01	15:00 – 16:30	All	Presentation Workshop II	R005 (bldg 30.28)	Frau Jüttner
2026-06-08 – 2026-06-19		Individual	Presentations "State-of-the-Art"Einzeln: 20 Min Vortrag + 10 Min Fragen 2er-Gruppe: 25 Min Vortrag + 12 Min Fragen 3/4er-Gruppe: 30 Min Vortrag + 15 Min Fragen		(students)
2026-06-08 – 2026-07-31		Individual	Seminar Paper Submission	n/a	students	after State-of-the-art presentation, deadline set by / coordinated with advisor
2026-06-25	14:00 – 16:00	Lecture	Erkenntnistheorie	SR 010 (bldg 50.34)	Prof. Snelting	In German
2026-07-02	14:00 – 15:30	Lecture	Research Question	t.b.a.	M. Ulbrich
2026-07-07	09:45 – 11:15	Lecture	Theory of Science	SR 010 (bldg 50.34)	Prof. Reussner
2026-07-09	14:00 – 15:30	Lecture	Writing Proposals	SR 010 (bldg 50.34)	M. Ulbrich
2026-07-16	14:00 – 15:30	Lecture	Experiment Design	SR010 (bldg 50.34)	Hr. Bechberger
2026-07-20 – 2026-07-31		Individual	Short Project Presentation / Mandatory Meeting With ExaminerHave an update meeting with your examiner or present the current state to the research group. Suggested lengths for talks: One-Person Project: 5 Min Talk + 5 Min Discussion Two-Person Project: 8 Min Talk + 8 Min Discussion Three-/Four-Person Project: 10 Min Talk + 10 Min Discussion		students
2026-09-16 – 2026-09-30		individual	Project Presentation (Final Presentation Semester I)		students	before submission of project proposal
until 2026-09-30		Individual	Project Proposal	n/a	n/a	Individual timeline can be negotiated with resp. examiner.
until 2026-09-30		Individual	Exam First Semester		n/a	coordinated with examiner

Second semester

Date	Time	Event
Lecture week 1	1.5 hours	Kickoff & documentation of scientific progress
Lecture week 4	1.5 hours	Statistical analysis
Lecture week 5	2.0 hours	Writing a paper
Lecture week 9	5.0 hours	Writing workshop I - Models & techniques for scientific writing
Lecture week 11	6.0 hours	Writing workshop II - Writing abstracts & precise formulation

As part of Praxis der Forschung there are additional presentation dates (three in the first semester, two in the second). In addition, the contents of each semester are assessed with an oral exam.
Further dates, details and possible changes can be found in ILIAS. Please register there.

Events and dates from past years can be found in the archive.