Transformer models are powerful for text reasoning and generation, but they can be unstable when the input text is slightly changed (e.g., paraphrasing, wording changes, noisy retrieval results). This thesis investigates Lipschitz-based robustness methods to improve the stability of transformer models on textual input, with application cases such as neural-logical inference and retrieval-augmented generation (RAG).

Small perturbations in text input or retrieved context can lead to large changes in model output, even when the meaning remains similar. This limits reliability in reasoning and RAG systems. The thesis aims to study and evaluate methods that enforce or approximate Lipschitz continuity (or related stability properties) in transformer-based models, and to measure whether they improve robustness in practice.

For Bechalor's thesis (preferred):

• Background in machine learning / deep learning
• Basic knowledge of NLP and transformer models
• Programming skills in Python (preferably PyTorch)
• Interest in robustness, trustworthy AI, or formal reasoning

For Master’s thesis (preferred):

• Stronger experimental skills and ability to read research papers
• Interest in extending methods or conducting deeper evaluations