Jakob Sturm

M. Sc. (he/him) (external PhD student with BMW)

Research interest

The focus of my research is on Retrieval-Augmented Generation (RAG) as an approach for domain adaptation. While pretrained large language models (LLMs) offer great potential, they need to be adapted to handle domain-specific tasks and incorporate relevant knowledge. RAG addresses this need, but it is not yet a perfect solution. I aim to contribute to the identification and resolution of existing challenges.

Current research topics include:

• Evaluation methods for RAG, both end-to-end and component-wise
• Domain adaptation strategies for the retrieval component of RAG
• Comparison and integration of RAG with fine-tuning approaches
• Identification and analysis of key pain points in RAG systems

Supervision of Theses

Important: Currently, I am not available to supervise any additional topics, so please only reach out if you are interested in one of the listed open projects!

If you are looking for a thesis or guided research topic and are motivated to work on a project related to my research interests, don't hesitate to contact me via email! Please attach your CV, Transcript of Records, and a short (< 400 words) introduction about yourself and your motivation. As I only have very limited capacities, please understand that I cannot offer a thesis to every interested student. However, I will try to answer your enquiries nevertheless, i.e. if you haven't heard from me, I am still deciding whether I can provide you a topic or not.

(M = master thesis; B = bachelor thesis) (TUM = Internal TUM thesis; BMW = external thesis, linked to a thesis position at BMW)

Open Topics

• (M) (TUM) Assessing the potential of hybrid retrieval methods for special domain items with focus on large text elements.
• (M) (TUM) Comparing RAG and Finetuning with respect to output quality and costs
• (M) (TUM) Assessing the potential of embedding model blackbox finetuning for domain adaptation

Currently ongoing Thesises

• (M) S3EK: boosting task-Specific Semantic Search with Expert Knowledge data
• (B) Assessing the role of generator and retriever within a RAG pipeline on the Question Answering task with the help of NLP metrics
• (M) Agentic RAG, relevance feedback and keyword identification from search histories

Finished Thesises

• (2025) (M) Advanced Methods for Finding Related Tickets Based on Semantic Search (Continous pretraing and finetuning for IR; hybrid retrieval)
• (2024) (M) Evaluation of Retrieval augmented Generation architectures
• (2024) (M) Boosting Quality Control in the Automotive Industry using LLMs and Contrastive Learning (Blackbox finetuning for Embedding Models)