Hi there,

my name is Luca, and I’m a PhD student at the University of Bern. Currently, my research focuses on Natural Language Processing (NLP) and Black-Box Optimization (BBO). I am part of a four-year project funded by the Swiss National Science Foundation, which aims to detect and quantify the use of sustainable criteria in public procurement processes in Switzerland. Lately, I have been working on Information Retrieval, Legal Summarization, and the application of BBO methods to NLP problems.

If you are interested in my work, feel free to explore my publications and activities, or contact me via email.

Latest publications

See all

Example of two parameter blocks with different types of parameters and different optimizers within the MixDiBB framework.

MixDiBB: Distributed Black-Box Optimization for Mixed-Parameter Search Spaces

Luca Rolshoven, Matthias Stürmer, Giuseppe Cuccu (2026)

Proceedings of the Genetic and Evolutionary Computation Conference — New York, NY, USA

We introduce MixDiBB, a framework for large-scale mixed-parameter black-box optimization. MixDiBB works by leveraging partial separability, optimizing highly correlated parameter subsets asynchronously on a distributed cluster or multiple CPUs on a single machine. We show that MixDiBB outperforms state-of-the-art mixed-parameter optimizers on both low-dimensional and high-dimensional problems with 12,000 parameters.

Paper Code

Grouped bar chart showing two fine-tuned LLMs of the Qwen2.5 family and two frontier models evaluated on the SLDS test set across ROUGE, BLEU and LLM-as-a-Judge metrics.

Unlocking Legal Knowledge: A Multilingual Dataset for Judicial Summarization in Switzerland

Luca Rolshoven, Vishvaksenan Rasiah, Srinanda Brügger Bose, Sarah Hostettler, Lara Burkhalter, Matthias Stürmer, Joel Niklaus (2025)

Findings of the Association for Computational Linguistics: EMNLP 2025 — Suzhou, China

We present SLDS, a cross-lingual dataset of 20K Swiss court rulings with expert-authored headnotes in German, French, and Italian. Evaluation results suggest that factual accuracy matters more than reasoning for legal headnote generation.

Paper Code Dataset Models

Identifying Sustainability in Public Tendering

Luca Rolshoven, Veton Matoshi, Tilia Ellendorff, Sarah Hostettler, Rahel Meili, Judith Binder, Matthias Stürmer (2025)

Proceedings of the 2nd Workshop on Green-Aware Artificial Intelligence (Green-Aware AI@ECAI 2025) — Bologna, Italy

We propose a lightweight NLP pipeline for identifying sustainability criteria in Swiss public procurement documents by computing cosine similarity between sentence embeddings and official Sustainable Procurement Criteria catalogs. A threshold-based evaluation, validated by human experts and an LLM-as-a-Judge ensemble, shows that Gemini 2.0 Flash achieves substantial agreement with expert judgments, offering a scalable alternative to manual annotation.

Paper

Zero-Shot Award Criteria extraction via Large Language Models from German Procurement Data from Switzerland

Veton Matoshi, Luca Rolshoven, Matthias Stürmer (2024)

Proceedings of the 9th edition of the Swiss Text Analytics Conference — Chur, Switzerland

We show how a single zero-shot German LLM prompt can automatically extract and structure Award Criteria from Swiss calls for tenders. By jointly performing classification, named entity recognition, relation extraction, and formatting, our approach enables scalable monitoring of sustainability-related procurement practices. We evaluate on 167 annotated tenders and release both code and data publicly.

Paper Code

Illustration of the core architecture of DiBB with a head node sending and receieving the global mean, whereas several blockworkers evaluate their local candidate solutions as part of the global mean.

DiBB: Distributing Black-Box Optimization

Giuseppe Cuccu, Luca Rolshoven, Fabien Vorpe, Philippe Cudré-Mauroux, Tobias Glasmachers (2022)

Proceedings of the Genetic and Evolutionary Computation Conference — New York, NY, USA

We introduce DiBB, a framework for scalable black-box optimization that decomposes parameters into correlated blocks and optimizes them in parallel, preserving key properties of the underlying algorithms while enabling efficient distributed execution and large-scale applications.