Repository for Publications and Research Data
News from the ETH Library
Recently Added
The Influence of Different Types of Trust, Perceived Objectivity, and Perceived Risk on the Preferred Involvement of AI for Subjective and Objective Tasks
Item type: Journal Article
Michel, Fabienne; Siegrist, Michael (2026)
Artificial intelligence (AI) is increasingly involved in decision-making. While many consumers currently use AI for minor tasks such as music recommendations or for improving writing, it can also be used for more consequential decisions such as medical diagnoses or financial recommendations. A declaration of AI content is not common, and consumers may not always be aware that AI is involved in the decision-making. Therefore, it is important to understand their perception of AI for different tasks and how different kinds of trust influence the degree to which AI involvement is accepted. The psychometric paradigm was adapted to analyze participants' acceptance, perceived objectivity, and perceived risks in the event of errors for a list of different tasks. To analyze the role of trust, measures for general confidence, general trust, and social trust, as well as AI-trust and AI-distrust were collected. The online survey was completed by 923 Swiss participants. The results show that participants' preferred levels of AI involvement were influenced positively by the degree of subjectivity and negatively by the perceived risk associated with the task. Preferred AI involvement was highest for "Searching for information," "Making a weather forecast," and "Programming software" but lowest for "Delivering a court judgment," "Selecting a candidate for a job," and "Driving a car." General confidence, general trust, and social trust indirectly influenced acceptance of AI through AI-trust, but not AI-distrust. The findings indicate that although specific AI-trust has the largest influence on acceptance of AI, general types of trust are also important for evaluating AI in decision-making.
Benchmarking Positional Encodings for GNNs and Graph Transformers
Item type: Conference Paper
Grötschla, Florian; Xie, Jiaqing; Wattenhofer, Roger (2026)
Positional Encodings (PEs) are essential for injecting structural information into Graph Neural Networks (GNNs), particularly Graph Transformers, yet their empirical impact remains insufficiently understood. We introduce a unified benchmarking framework that decouples PEs from architectural choices, enabling a fair comparison across 8 GNN and Transformer models, 9 PEs, and 10 synthetic and real-world datasets. Across more than 500 model-PE-dataset configurations, we find that commonly used expressiveness proxies, including Weisfeiler-Lehman distinguishability, do not reliably predict downstream performance. In particular, highly expressive PEs frequently fail to improve, and can even degrade performance on real-world tasks. At the same time, we identify several simple and previously overlooked model-PE combinations that match or outperform recent state-of-the-art methods. Our results demonstrate the strong task-dependence of PEs and underscore the need for empirical validation beyond theoretical expressiveness. To support reproducible research, we release an open-source benchmarking framework for evaluating PEs for graph learning tasks.
Structure, Geometry and Noise in Stochastic First-Order Optimization
Item type: Doctoral Thesis
Fatkhullin, Ilyas (2026)
Modern optimization tackles problems far beyond the idealized regimes of textbook theory. Simple first-order methods often remain reliable despite non-convex landscapes, complex constraints, non-Euclidean geometries, and occasional extreme observations. To explain this reliability, this dissertation focuses on three interrelated themes.
The first concerns landscape structure. Some problems appear non-convex only because they are written in inconvenient variables: after an unknown or implicit change of variables, they become convex. Such formulations arise in models from control and reinforcement learning and in applications from revenue and inventory management; across these problems, natural decision variables often obscure a convex reformulation. The thesis shows that stochastic subgradient methods can exploit this hidden convexity without ever using the convexifying map, proving global convergence guarantees. The thesis then carries this perspective to functional constraints, where guarantees must certify not only objective value but also feasibility.
The second theme is geometry. Mirror descent is a classical first-order method that changes the local model used by the algorithm. This thesis develops a general convergence theory of non-convex stochastic mirror descent for general Bregman divergences, without requiring large mini- batches or globally Lipschitz gradients of the distance-generating function. Under relative smoothness and bounded variance, a new Lyapunov anal- ysis controls a strong stationarity certificate that directly measures the decrease available to the mirror step. This framework leads to applications in differentially private learning and variational inference.
The third theme is noise. Classical SGD theory typically assumes bounded second moments, yet stochastic gradient noise in many modern applications is often heavy-tailed. Under only finite p-th moment assumptions, we prove that SGD remains a sharp and meaningful baseline in expectation: it achieves optimal rates in convex settings and tight rates specific to SGD in non-convex settings. At the same time, our lower bounds show that these expectation guarantees do not translate into comparably strong high- probability convergence, clarifying why popular adaptive methods often outperform vanilla SGD.
Together, these results clarify the capabilities and limitations of stochastic first-order methods and inform the design of more efficient algorithms.
Premium Subsidies and Healthcare Use under Universal Coverage
Item type: Working Paper
Anderes, Marc (2026)
I study how changes in premium subsidies affect healthcare use within Switzerland’s mandatory insurance system. Using administrative data for 93,631 households in 2023–2024, I instrument actual subsidies with cantonal subsidy schedules. An additional CHF 100 of annual subsidy raises annual gross covered healthcare spending by approximately CHF 35. The response is concentrated in ambulatory care and pharmacy dispensing, including increased physician-delivered psychiatric care, while the share of households with any covered claim changes little. Greater generosity also increases switching toward unrestricted provider choice, without detectable systematic deductible adjustment or improvement in recorded payment difficulties. The financing burden of mandatory insurance thus affects how households use existing coverage, even without direct changes in contractual cost sharing.
LLM-Driven Architectural Geometry Generation: From Prompt to Parametric Model
Item type: Conference Paper
Liu, Minjuan; Tai, Hung-Chen; Shen, Chen; et al. (2026)
Traditional architectural geometry design is a critical yet labour-intensive stage in the design process. The emergence of generative artificial intelligence (GenAI) models, such as diffusion models, has enabled rapid explorations of architectural conceptual images and 3D massing. But their limited semantic understanding, low editability, and lack of interactive refinement constrain the professional use. Large language models (LLMs), with their capacity for semantic reasoning, multimodal integration, and procedural script generation, offer new opportunities to address these limitations. This study investigates how LLMs can support early-stage geometry generation through an experiment with 27 graduate students, producing over 270 design alternatives for an office building in Beijing. Three key findings emerged: 1) Compared to diffusion-based platforms, LLMs can generate 3D models with clearer geometric boundaries, higher editability, multi-round refinement, and stronger downstream integration. 2) LLMs can uniquely generate parametric models that yield multiple design alternatives, demonstrating an ability to encode design logic and thereby offering a higher level of controllability than conventional 3D meshes. 3) We identified effective strategies to improve efficiency, including explicitly articulating design logic, providing reference models or images, and facilitating cross-model collaboration. Our findings establish LLMs as powerful co-creative partners, advancing the potential of AI-assisted architectural design.