Master’s theses: Creation of a Spatial Reasoning Benchmark

22.05.2026, Diplomarbeiten, Bachelor- und Masterarbeiten

This thesis aims to develop a benchmark for evaluating the spatial reasoning of LLMs. Since AI models often struggle with 3D planning, LEGO bricks serve as a discretized medium to quantify this intelligence. Key tasks include generating synthetic datasets, defining metrics for spatial correctness, and validating AI-generated blueprints within 3D simulations.

Current Large Language Models (LLMs) demonstrate impressive capabilities in text processing but often reach their limits when faced with complex spatial reasoning tasks. While logical reasoning in language has advanced, the ability to precisely plan and manipulate 3D structures remains a significant challenge. LEGO bricks provide an ideal, discretized medium to make the spatial intelligence of AI models measurable. Our goal is to develop a benchmark (similar to approaches like Minebench) that evaluates the extent to which LLMs are capable of understanding assembly instructions, designing stable 3D structures, or correctly describing spatial relationships between objects.

The following fields of research are open to you for your work in the context of spatial reasoning and benchmark development:

• Development of methods for the synthetic generation of LEGO datasets
• Definition of metrics to evaluate the spatial correctness of AI-generated blueprints
• Investigation of reasoning strategies (e.g., Chain-of-Thought) for solving 3D construction tasks
• Integration of LLMs into 3D simulators to validate assembly instructions

We offer:

• An innovative research topic at the intersection of LLMs and geometry
• An exciting and future-oriented research environment
• Writing in German or English possible

Requirements:

• Initiative and a strong passion for problem-solving and spatial logic
• Advanced knowledge of Python (ideally experience with transformers, openAI/langchain libraries)
• Experiences with bricks and the file format ldr are additional advantages
• Basic knowledge of machine learning and working with LLMs

The theses can start immediately. Are you interested in this topic or have questions?
Feel free to contact André Schamschurko ([andre.schamschurko@tum.de])

Kontakt: andre.schamschurko@tum.de