• 09:15 - 09:30 CEST
• Thomas Schmid
• tbd.

Opening & Welcome

• 09:30 - 10:30 CEST
• Martin Hebart
• tbd.

Keynote "Similarities and Differences of Representations in Humans and Artificial Intelligence"

Recent AI models have been shown to match or surpass human performance on many benchmarks, yet their success often tells us surprisingly little about the representations that support it. When two systems produce similar outputs, their decisions can be based on fundamentally different dimensions of the world. This leads to a central challenge for human-AI alignment: how can we identify what it is that leads to similar representations in humans and AI? In this talk, I will present work from our lab that uses tools from cognitive science to reveal the representational structure of human and artificial systems. Starting from large-scale human similarity judgments, I will show how object representations can be decomposed into interpretable dimensions that capture much of human perceptual and conceptual structure in representations. I will then describe how the same logic can be applied to deep neural networks, revealing striking differences between how humans and artificial models make sense of the world. I will end by extending this approach across large model families to reveal what are universal dimensions of representations learned across models and how they relate to biological vision. Together, these findings highlight that applying the toolkit of cognitive science to the study of artificial systems can yield important similarities and differences in learned representations across biological and artificial neural systems.

• 10:00 - 10:30 CEST

Coffee Break

• 11:00 - 11:30 CEST
• Melle Mendikowski
• tba.

Bridging Computer Science, Economics, and Social Science: Reproducible LLM Workflows for Large-Scale Qualitative Text Coding

Large Language Models (LLMs) are creating new opportunities to scale qualitative research, but reliable and reproducible methods for integrating them into workflows are still emerging. This interdisciplinary project brings together researchers from computer science, economics, and the social sciences to design, implement, and evaluate LLM-assisted approaches for large-scale text classification and qualitative coding. Drawing on corpora of approximately 6,000 computer science papers and 544 research abstracts, we examine how LLMs can help identify research narratives, including portrayals of AI as augmenting or replacing human labor, alongside more complex qualitative coding tasks. We present QualCoder-LLM, an open-source Python pipeline that supports prompt experimentation, validation, coding, logging, and evidence-line verification while preserving human oversight and reproducibility. Our findings offer practical guidance for researchers seeking to use LLMs as scalable collaborators in systematic literature reviews and other text-intensive interdisciplinary research.

• 11:30 - 12:00 CEST
• Tracy Arndt
• tbd.

Practical Approaches to Large-Scale Linked Open Data

Over the past sixteen years, the German National Library has developed its Linked Open Data (LOD) services into a large‑scale, community‑oriented ecosystem. Our work spans three interconnected roles: data producer of authority and bibliographic records, data provider offering over 1.4 billion triples via open interfaces, and community member collaborating with researchers, developers, and cultural heritage institutions. This talk highlights how openly developed vocabularies, transparent workflows, and a freely accessible SPARQL endpoint support sustainable LOD infrastructures. We share insights from operating a high‑volume service, including challenges in data quality, update cycles, and increasing pressure from AI-driven scraping.

• 12:00 - 13:00 CEST

Lunch Break

• 13:00 - 13:30 CEST
• Paul-Niklas Kandora
• tba.

LinearizeLLM: An Agent-Based Framework for LLM-Driven Exact Linear Reformulation of Nonlinear Optimization Problems

Reformulating nonlinear optimization problems into solver-ready linear or mixed-integer linear problems is a common but expertise-intensive task in mathematical programming. We study this task as a verification-aware reformulation problem: given a raw LaTeX optimization problem, can a workflow recover the reformulation-relevant problem structure, apply known exact recipes for a fixed supported reformulation family, and validate returned candidates by execution? We propose LinearizeLLM, a multi-stage language-model-supported workflow that detects supported nonlinear patterns, records their nesting structure, applies a deterministic bottom-up reformulation order, and generates solver-ready LP/MILP candidate formulations with reformulation records. The current implementation supports absolute values, minimum and maximum operators, binary-continuous products, linear fractional terms under sign conditions, and monotone objective transformations. On ComplexORNL, a 40-instance controlled benchmark derived from ComplexOR and synthetic nested models, LinearizeLLM has higher observed end-to-end objective-preservation success than the evaluated same-model one-shot reformulator and the raw-LaTeX-to-GDP-to-MILP pipeline. In order to address correctness beyond objective agreement, we add a solution-mapping verification step: 88 of 89 successfully-reformulated LinearizeLLM runs are mapped back to original decision-variable assignments that pass feasibility and optimality checks in the original reference problem.

• 13:30 - 14:00 CEST
• Andre Opris
• tba.

Understanding the Population Dynamics of NSGA-III via Runtime Analysis

NSGA-III is a widely used algorithm in evolutionary many-objective optimization and is designed for problems with more than three objectives, distinguishing it from the classical NSGA-II. Despite its practical success, the theoretical understanding of when and why NSGA-III performs well remains limited. In this talk, we help bridge this gap by presenting rigorous runtime analyses of NSGA-III on several classical benchmark problems with arbitrary numbers of objectives. Our results improve previous analyses in regimes where the population size asymptotically exceeds the size of the Pareto front. A key ingredient of our work is a detailed understanding of the population dynamics of NSGA-III. Notably, in the bi-objective setting, the obtained upper runtime bounds are asymptotically tighter than the best known bounds for NSGA-II. For suitable population sizes, NSGA-III even achieves a smaller expected runtime than NSGA-II. Finally, we complement these results by establishing tight runtime bounds for NSGA-III.

• 14:00 - 13:30 CEST
• John A. Warwicker
• tba.

Theoretical and Empirical Analysis of Classification Schemes within Neuroevolution and Evolutionary Neural Architecture Search

The choice of architecture of a neural network (NN) is crucial to its performance, and is usually found by a complex trial-and-error approach with significant manual input. Evolutionary methods have successfully been employed to automate this approach, leading to the promising field of evolutionary neural architecture search (ENAS). Despite its prevalence, the performance of ENAS is not well understood. Recent theoretical works analysing ENAS and the use of evolutionary approaches for parameter training within NNs have appeared; however, these approaches either assume a pre-defined architecture or an oracle-based training approach. In this work, we present the first theoretical analyses of ENAS with in-built parameter adaptation for binary classification problems. We firstly analyse ENAS with the oracle-based \textit{Neuron Parameter Training Strategy} for the benchmark \textsc{Uniform} problem class, by isolating the performance for different mutations to the architecture, and comparing local and global training strategies. We show that the oracle-based approach is inefficient as it relies on large jumps or a random walk in the parameter space, and thus has an expected runtime complexity that increases with the problem size and the chosen resolution. We then analyse the training approach without an oracle, and showcase that it can be more efficient as it relies on hillclimbing in the parameter space. Nevertheless, to overcome inefficiencies associated with poor initialisations, we highlight the need for restart strategies or alternative classification schemes.

• 14:30 - 15:00 CEST

Coffee Break

• 15:00 - 15:30 CEST
• Jixiang Qing
• tba.

Bayesian Optimization over Graphs via Global Acquisition Optimization

Bayesian optimization (BO) is highly effective for tasks like molecular design and neural architecture search (NAS). However, optimizing acquisition functions in discrete, combinatorial graph spaces, especially under structural constraints, remains a significant challenge. This talk addresses this critical bottleneck by proposing a novel approach for global acquisition optimization. We demonstrate how to encode both the graph search space and shortest-path graph kernels using mixed-integer programming (MIP). This formulation enables the exact, global optimization of acquisition functions such as the Lower Confidence Bound (LCB). We demonstrate the effectiveness of this scheme on standard molecular design and NAS benchmarks.

• 15:30 - 16:00 CEST
• Maren März
• tba.

From Learning Curves to Licensing Outcomes: An Analysis of Progress Test Data

Progress tests provide longitudinal insight into students’ knowledge development and have demonstrated predictive value for licensing examinations. Building on an earlier feasibility study using individual‑level LSTM models to forecast state licensing exam outcomes (M2) from item‑level behaviour, response patterns, and performance clusters, we extend the analysis to the cohort level. Using aggregated progress test means from three faculties, we extracted linear and non‑linear learning‑curve features and modelled their association with cohort‑level deviations from the national M2 mean using Elastic Net, Bayesian Ridge, and XGBoost. Individual predictions exceeded field‑comparison accuracy from the 6th semester onward. At the cohort level, steeper and higher learning curves—particularly, but not limited to, in later semesters—were consistently associated with better M2 outcomes. Together, both modelling strands indicate that the level and shape of progress‑test learning trajectories carry substantial predictive information for licensing performance and could be used to provide feedback at individual and institutional levels.

• 16:00 - 16:30 CEST
• Christian Zinke-Wehlmann
• tba.

From Ambition to Reality: Lessons Learned from the KISS Project

How can AI help manufacturing networks respond faster, more transparently, and more resiliently in times of crisis? The KISS project set out to explore this question through the idea of an AI-based semantic networking platform for rapid supply networks, connecting platform technologies, additive manufacturing, and artificial intelligence. This talk offers a reflective recap of the project journey: what worked, what proved harder than expected, and what we learned about translating semantic AI concepts into practical, trustworthy, and organisationally usable infrastructures. The talk highlights the pitfalls and most promising outcomes, such as an agentic approach.