Back to the main page of "Practice of Research"

Winter Semester 2026/27

Project Group "Practice of Research"

Announcements:

Slides containing general information on Practice of Research (from last year)

Research Groups Offering Topics in Winter Semester 2026/27

The research groups participating in Winter Semester 2026/27 are:

The following list provides an overview of the topics offered this semester. If you are interested in a topic or have questions about it, you can contact the relevant staff member. This list will be expanded. The final list of offered topics will be announced at the information event and topic presentation at the beginning of the semester (see above) and can subsequently be found in the published slides.

Note: Experience has shown that it is worthwhile to contact the relevant staff members in advance in order to avoid disappointment at the topic presentation. If you are interested in carrying out a project as part of Practice of Research, but your preferred topic is not listed here, then in exceptional cases additional topics may still be submitted until the end of the first week of lectures. 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 interested).

List of Projects for Winter Semester 2026/27

Last update: 27 October 26

Offered Available Projects

#	Project	Research Group(s)	Supervising Staff
3.	GridCascade: A Synthetic Benchmark for Spatio-Temporal Learning of Cascading FailuresThe graph machine learning community increasingly recognizes that progress is held back by a lack of real or realistic synthetic benchmarks. This limits both the development and rigorous evaluation of new methods, as well as their transfer to real-world applications. Power grids are particularly suited for building such benchmarks: a single line failure can overload other lines and trigger a cascade that spreads across the grid. Such cascades change the graph's topology, edge flows and node states abruptly and at irregular times, properties that current benchmarks rarely test together. As a first step toward GridCascade, a synthetic benchmark of cascading failures in power grids, this project surveys quasi-steady-state and dynamic cascade simulators, maps the data they provide to the requirements of state-of-the-art spatio-temporal graph learning methods, and creates a first prototype dataset based on the selected simulator. GridCascade should record the full sequence of failures, providing the grid state and topology at each stage of the cascade. Benchmark tasks should include predicting the next failures, the propagation path, and the final extent of a cascade. Since the simulation reveals which failures cause subsequent ones, GridCascade should also provide ground-truth explanations for evaluating explainability methods. First state-of-the-art spatio-temporal graph models will be tested on the prototype, and their limitations will point to directions for new methods. The long-term goal is to release the dataset and code openly to support the spatio-temporal graph learning community.	IAI Schäfer	Martin Sadric
4.	Designing and Evaluating Robot-Supported Reading Activities in Primary EducationThis project investigates how social robots can support children's learning to read in primary education. Social robots have shown potential as educational technologies, yet important questions remain regarding how they can support learning in real classrooms, how their use evolves over time, and how they are experienced by teachers and children in educational settings. Building on prototype development and a long-term collaboration with a local primary school, students will design, deploy, and evaluate robot-supported reading activities. In doing so, they will investigate how child-robot interactions can support reading acquisition, how reading activities can be refined through classroom use and teacher feedback, and how social robots can become part of educational practice. Students will contribute to the development and evaluation of research prototypes through interaction design, classroom deployments, user studies, and empirical research. Good German language skills are required.	IAR Bruno	Utku Norman Irina Rudenko
5.	Benchmarking the AI-Content Provenance Ecosystem under the EU AI ActSince August 2026, Article 50 of the EU AI Act requires providers to mark AI-generated images, video and audio in a machine-readable way. Yet provenance is a property of the whole ecosystem: models, providers that add markings, and social media and messaging platforms that display or strip them. Existing audits are one-off, manual, small-scale snapshots of single links. The goal is to design and develop a benchmark infrastructure, run by AI computer-use agents, that runs periodically: it generates content with major providers, shares it via social media and messaging, records labels and re-verifies files, also under everyday handling and adversarial attacks. Its results will inform policymakers and civil society, holding Big Tech accountable. This project requires solid software-engineering experience and interest in AI policy. Bemerkung: I won't be able to be physically present at the seminar opening meeting. I'm happy to either join virtually, or another member of the team will present the project.	TM Strufe	Shlomi Hod
6.	Participatory Design and Evaluation of a Social Robot for Supporting Self-Regulated Learning in After-School Homework SessionsThis thesis project explores the participatory design and evaluation of a social robot to support self-regulated learning during after-school homework sessions at Markgrafen Gymnasium in Durlach. The robot aims to assist teachers in monitoring attendance and homework completion while encouraging younger students to work independently. The project involves refining an existing prototype based on the needs of students and teachers, designing effective human-robot interactions, and deploying the system in a real school environment. Through classroom studies, qualitative and quantitative data collection, and stakeholder feedback, the research will investigate design requirements, user experiences, and the robot’s potential to foster students’ autonomy and self-regulation during homework sessions.	IAR Bruno	Romain Maure Irina Rudenko Utku Norman
7.	Perceived AI Ownership in Collaborative Decisions in Virtual EnvironmentsThis project investigates how perceived ownership of AI in the team affects the negotiation and information sharing behavior and in turn the task performance. The project implements an experimental setup to test these effects in Virtual reality using unity. It also implements an LLM-based AI assistance. A focus lies on the large-scale evaluation of the experiment, the grounded theoretical work for explaining and anticipating user behavior in this situation, and a clear and detailed description of arguments of expected behavior. Part of the VR setup is already implemented, but adoptions, refinements and optimization of the audio conversation capabilities are necessary. Furthermore several pre-tests need to be conducted before the experiment can be tested in a large scale study in a VR lab. The supervisors will cover participation fees for participants, in line with standard practices for the KD²Lab pool. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAR Bruno	Barbara Bruno Jella Pfeiffer
9.	Bayesian Knowledge Distillation for Motion Prediction ModelsMotion prediction is a key component of autonomous driving systems, enabling the anticipation of future behavior of surrounding traffic participants. Recent advances have led to highly capable multi agent motion prediction models based on transformers and graph neural networks that achieve state of-the-art performance on large-scale benchmarks. However, these models are often computationally expensive and difficult to deploy in latency-critical applications. Knowledge Distillation (KD) aims to transfer knowledge from a large, high-performing teacher model to a smaller and more efficient student model. Classical distillation methods primarily focus on matching predictions and retaining the accuracy of the teacher model. However, a major open research question is whether distilled models can also retain the uncertainty characteristics and robustness properties of their teachers. In particular, it remains unclear how uncertainty-aware distillation affects performance in safety-critical long-tail scenarios and under out-of-distribution (OOD) conditions, which are often underrepresented in standard benchmark evaluations but highly relevant for real-world deployment. Bayesian distillation methods are particularly attractive in this context, because understanding prediction confidence is crucial for safe decision-making in autonomous driving. The goal of this thesis is to investigate, implement, and evaluate Bayesian distillation methods for autonomous driving motion prediction. The work can be conducted using an existing research software stack in Python, Pytorch containing state-of-the-art motion prediction models, large scale datasets, and evaluation pipelines for benchmark, long-tail, and OOD performance assessment. Bemerkung: Research Group: Prof Nadja Klein (MBD, SCC)	MBD Klein	Christian Schlauch
10.	Conformal Motion Prediction for Autonomous DrivingMotion prediction is a key component of autonomous driving systems, enabling the anticipation of future behavior of surrounding traffic participants. Recent advances have led to highly capable multi agent motion prediction models based on transformers and graph neural networks that achieve state of-the-art performance on large-scale benchmarks. However, these models typically rely on heuristic confidence scores and evaluation procedures based on a fixed number of predictions (e.g., Top-K trajectories). When confidence estimates are not well calibrated, they provide limited guarantees regarding the coverage of plausible future behaviors. This can be particularly problematic in safety critical and ambiguous traffic scenarios where uncertainty-aware decision making is essential. Conformal prediction has recently emerged as a powerful framework for uncertainty quantification with finite-sample coverage guarantees. Conformal prediction uses calibration data to construct prediction sets that contain the true outcome with a user-defined probability under mild assumptions. While conformal methods have been successfully applied to classification and regression tasks, their application to multi-modal motion prediction remains largely unexplored. The goal of this thesis is to investigate conformal prediction methods for autonomous driving motion prediction. In particular, the thesis will explore how trajectory hypotheses generated by state-of-the art motion prediction models can be transformed into calibrated trajectory sets with statistical cov erage guarantees while maintaining high prediction quality and diversity. The work can be conducted using an existing research software stack in Python and Pytorch containing state-of-the-art motion prediction models, large-scale datasets, and evaluation pipelines for benchmark, long-tail, and OOD performance assessment. Bemerkung: Research Group: Prof. Nadja Klein (MBD, SCC)	MBD Klein	Christian Schlauch
11.	Sampling-Based Model Predictive Control for Motion PlanningModel Predictive Control (MPC) is a widely used framework for motion planning and decision making in autonomous driving. Classical MPC approaches typically formulate motion planning as a nonlinear optimization problem subject to vehicle dynamics, safety constraints, and prediction models of surrounding traffic participants. While optimization-based formulations can provide structured solutions and enable the incorporation of domain knowledge, they often become computationally challenging in highly nonlinear or multi-modal planning scenarios. Recent advances in robotics and model-based reinforcement learning have demonstrated the effectiveness of sampling-based MPC methods such as the Cross-Entropy Method (CEM), and Model Predictive Path Integral (MPPI) control. Instead of solving an optimization problem using gradients or numerical solvers, these methods generate and evaluate large numbers of candidate trajectories through stochastic sampling. Sampling-based methods naturally support non-convex objectives, black-box models, and parallel execution, making them attractive for the application on hardware accelerators. However, a key limitation of existing approaches is that they typically assume unimodal Gaussian distributions over control inputs. Motion planning and behavior prediction in autonomous driving are inherently multi-modal: multiple plausible maneuvers often exist simultaneously, such as changing lanes to the left or right, yielding at an intersection, or overtaking a slower vehicle. Maintaining these discrete maneuver hypotheses within a sampling-based MPC framework remains an open research challenge. The goal of this thesis is to investigate, implement, and evaluate sampling-based MPC algorithms for autonomous driving and to extend these approaches using mixture-based uncertainty models that represent multiple maneuver hypotheses. The work can be conducted using an existing research software stack in Python, PyTorch and CasADI, containing baseline MPC implementations, a simulation environment with NuPlan scenarios, and evaluation pipelines for benchmark, long-tail, and OOD performance assessment. Bemerkung: Research Group: Prof. Nadja Klein (MBD, SCC)	MBD Klein	Christian Schlauch
17.	Classical Probabilistic ArgumentationThis research project develops a probabilistic extension of classical argumentation theory. Debates are modelled as collections of logical arguments that restrict which combinations of claims can consistently be accepted. By assigning probabilities to the remaining logically possible worlds, the framework represents degrees of belief and uncertainty while preserving the classical structure of arguments and counterarguments. The project investigates how different prior beliefs can be represented, including through hidden statements and arguments; how beliefs change when new arguments or uncertain evidence are introduced; and how factor graphs and Markov-style independence relations can identify which claims are relevant to one another. It will relate the approach to existing work in probabilistic argumentation, Markov logic, and weighted model counting. The outcome will be formal definitions, theoretical and computational results, and a research report or paper draft.	KASTEL Beckert	Mattias Ulbrich Gregor Betz

Assigned Projects

Participants for these projects have already been found in advance.

#	Project	Research Group(s)	Supervising Staff
1.	Understanding Long-Range Dependencies in Hypergraph Neural NetworksGraph neural networks (GNNs) often struggle to capture long-range dependencies. Their limitations include oversquashing, where information from many nodes is compressed into fixed-size embeddings, and rank collapse, which reduces the diversity of node representations and can lead to oversmoothing. More generally, propagation dynamics can cause signals to vanish or grow uncontrollably across layers. This project investigates how these limitations arise in hypergraph neural networks within a topological deep learning framework. Unlike ordinary graphs, hypergraphs represent interactions among multiple nodes, raising questions about how higher-order connectivity affects information flow. The work will combine a review of existing approaches with theoretical analysis and controlled experiments on tasks requiring long-range propagation. It aims to identify the roles of hypergraph structure and model dynamics, and explore architectural principles that preserve information over long distances. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	IAR Neumann	Tai Hoang
2.	Input Models for Data ProfilingData profiling is the extraction of metadata from relational databases. This metadata includes multi-column dependencies like e.g. unique column combinations (UCCs) or functional dependencies (FDs). There are several algorithms known that discover all minimal UCCs/FDs of a database in time that is much faster than worst-case asymptotic analysis can currently explain. This project aims to design and analyze a more realistic input model for real-world databases that arise in data profiling. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	ITI Bläsius	Thomas Bläsius Martin Schirneck
8.	High-level Modeling of Quantum Software with Quantum Variables and HamiltoniansMost current models of quantum software are at the quantum circuit level. This PdF should investigate novel, high-level modelling abstractions of: (a) quantum variables, as incorporated in novel quantum programming languages; and (b) Hamiltonians, as abstractions that also apply to analogue quantum hardware. The integration of these new abstractions with quantum circuits through the treatment of composition as a first-class citizen should also be investigated. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	KASTEL Schaefer	Joshua Ammermann
12.	Advancing Quantum Circuit Editing with Noise Inspection, Debugging and correct-by-construction SubcircuitsA variety of tools for quantum circuit editing exist. This PdF should provide an overview of current editing methods and examine how a Quantum IDE can facilitate advanced circuit editing by: (a) modelling and inspecting the effect of noise in quantum circuits (e.g. through quantum depolarising channels); (b) advanced circuit inspection and debugging capabilities; and (c) integrating subcircuits created via the Quantum Correctness by Construction (QbC) approach. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	KASTEL Schaefer	Joshua Ammermann
13.	Traceability Link Recovery Meets Formal VerificationSee PDF document. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	KASTEL Koziolek	Dominik Fuchß
14.	Efficient uncertainty-aware consistency analysisModel-driven development of complex software-intensive or cyber-physical systems involves the creation of heterogeneous artifacts and models. Although most modeling tools seem to convey certainty for each value or model state, the information they represent often relies on assumptions and so, by necessity, include a degree of uncertainty. When specifying and managing such uncertainty within a single model, it is hard to determine its effect on the set of modeling artifacts and as well, on model consistency and correctness. The potential effects of uncertainty can be unclear, especially when models are automatically kept consistent. Effects become even more unclear when consistency and correctness are determined not solely by constraints but for a given model state, also by simulation or other black-box mechanisms. For such analysis, it is necessary to explore the induced possible models, propagate changes, and evaluate the consistency of multiple artifacts. In this project, the efficient consistency analysis in such heterogeneous settings should be analyzed, and approaches should be developed and implemented. Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	KASTEL Koziolek	Nathan Hagel Arne Lange
15.	Assessment of actual disclosure risks in medical data analysesAssessment of actual disclosure risks in medical data analyses Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	TM Strufe	Thorsten Strufe Patricia Guerra-Balboa
16.	Privacy-preserving Machine Learning beyond DP-SGDPrivacy-preserving Machine Learning beyond DP-SGD Bemerkung: Für das Projekt ist/sind bereits 1 (oder mehr) Stud. registriert.	TM Strufe	Thorsten Strufe Felix Morsbach

Events and Dates

The methodology events are expected to take place according to the following schedule. Participation in the first event is, among other things, a prerequisite for participating in Practice of Research; important organizational information will be provided there.

Information Event

See PdF main page

First Semester

Date	Time	Mode	Event	Location	Lecturer	Notes
2026-10-29	12:30 – 14:00	Open	Information Event + Topic PresentationOpen information event where the PdF procedures are explained, questions can be asked. There will be a session where the advisors present their PdF topics.	tba (prob Info bib)	Mattias Ulbrich
2026-11-09	09:45 – 11:15	All	Literature Search in Informatics	t.b.a.	Fr. Sielaff
2026-11-26	14:00 – 18:30	Groups	Presentation TrainingYou participate in one group event.	t.b.a.	Fr. Jüttner
2026-11-27	14:00 – 18:30	Groups	Presentation TrainingYou participate in one group event.	t.b.a.	Fr. Jüttner
2026-11-30	08:00 – 12:00	Workshop	Project management I	SR Forum (bldg. 30.95)	Markus Lorenz
2026-12-03	14:00 – 18:30	Groups	Presentation TrainingYou participate in one group event.	t.b.a.	Fr. Jüttner
2026-12-04	14:00 – 18:30	Groups	Presentation TrainingYou participate in one group event.	t.b.a.	Fr. Jüttner
2026-12-11	14:00 – 18:30	Groups	Presentation TrainingYou participate in one group event.	t.b.a.	Fr. Jüttner
2026-12-14 – 2026-12-22		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
2026-12-14 – 2027-02-19		Individual	Seminar paper submission		(students)	after State-of-the-art presentation, deadline set by / coordinated with advisor
2026-12-14	08:00 – 12:00	Workshop	Project management II	SR Forum (bldg. 30.95)	Markus Lorenz
2027-01-01 – 2027-01-01			WS: Presentations		N.N.	[to be scheduled]
2027-01-14	14:00 – 16:00	All	Erkenntnistheorie(*) Non-German alternative will be provided.	SR 010 (bldg 50.34)	Prof. G. Snelting	In German (*)
2027-01-22	14:00 – 17:15	All	Workshop Research Question	t.b.a.	Ulbrich
2027-01-28	14:00 – 15:30	All	Theory of Science	SR 010 (bldg. 50.34)	Prof. Reussner
2027-02-05	14:00 – 15:30	All	Writing Proposals	SR 010 (Geb. 50.34)	Ulbrich
2027-02-08 – 2027-02-19		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
2027-02-11	14:00 – 15:30	All	Experiments and Becnhmarking	SR010 (bldg. 50.34)	J. Bechberger
until 2027-03-31		Individual	Project Proposal	n/a	n/a	Individual timeline can be negotiated with resp. examiner.
until 2027-03-31		Individual	Exam First Semester		n/a	coordinated with examiner

Second Semester (Outline)

Week	Duration	Event
Lecture week 1	1.5 hours	Kick-off and documentation of scientific progress
Lecture week 4	1.5 hours	Statistical evaluation
Lecture week 5	2.0 hours	Writing papers
Lecture week 9	5.0 hours	Writing workshop I: models and techniques for academic writing
Lecture week 11	6.0 hours	Writing workshop II: writing abstracts and precise formulations

As part of Practice of Research, there are also presentation dates (three in the first semester and two in the second). In addition, the content of each semester is assessed in an oral examination.
Further dates, details, and possible changes can be found in ILIAS. Please enrol there.

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