Shortcut or Understanding? Diagnosing LLM Type Prediction in Conceptual Models

Authors: Syed Juned Ali, Zhuoxun Zheng, Dominik Bork

This paper investigates how natural language labels and structural context influence the accuracy of automated type prediction in conceptual models across different AI paradigms. The findings reveal that while semantic labels are the primary drivers of performance, both fine-tuned encoders (such as BERT) and prompting-based decoders rely heavily on surface-level lexical shortcuts rather than true structural reasoning. These results highlight the need for future architectures that incorporate pragmatic, requirements-driven context.

Read preprint

From Gut Feeling to Data-Driven Decisions: Exploring Digital Twin Supported Decision-Making for Real Estate Management

Authors: Marianne Schnellmann, Henderik A. Proper

This work identifies key challenges faced by real estate managers when making sustainability-related decisions for existing buildings. It further provides a forward-looking perspective by demonstrating how goal-oriented requirements for a Digital Twin can be systematically derived to support data-driven decision-making in real estate management.

Uncovering LLM's Capabilities in Model-based Question Answering for UML Class Diagrams

Authors: Manuel Mischak, Charlotte Verbruggen, Philip Langer, Dominik Bork

This paper presents an extension of the bigUML modeling tool with an LLM-based conversational interface and explores whether large language models can effectively support users by answering comprehension questions. The study provides insights into how representation format, context size, and LLM provider influence performance in model-based question answering, contributing to the advancement of accessible, AI-assisted modeling.

Read preprint