With the rise of Large Language Models (LLMs) across various domains, there is a growing need to incorporate complex spatiotemporal data, such as maps (e.g., OpenStreetMap), for tasks that involve navigation, orientation, and practical decision making. The challenge lies in efficiently encoding and passing this spatiotemporal information to LLMs to enhance their ability to answer practical questions while maintaining computational efficiency.
This includes exploring data encoding strategies, modifying model architectures, and evaluating their effectiveness in practical scenarios such as navigation and orientation.
Download as PDF
Student Target Groups:
- Students in ICE, Computer Science or Software Engineering.
Thesis Type:
- Master Project / Master Thesis
Goals and Tasks:
- Conduct a thorough literature review on spatiotemporal data representations and in context learning in LLMs.
- Develop methods to efficiently encode and integrate map data (e.g., OpenStreetMap) into LLMs for in-context learning.
- Evaluate the performance of these methods in practical tasks, such as navigation and orientation.
- Compare the computational efficiency and accuracy of different integration approaches.
- Present the results of your work and summarize the outcomes in a written report.
Required competences and knowledge:
- Strong understanding of neural networks and LLMs.
- Proficiency in programming languages such as Python.
- Familiarity with deep learning frameworks like PyTorch or TensorFlow.
Starting date:
Contact: