Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders
International Conference on Artificial Neural Networks 2026
I am a principal investigator at the University of Hamburg, where I co-lead the DFG projects LUMO and E4-MALM. My research focuses on trustworthy AI systems that can reason robustly, handle multimodal information (e.g., language, vision, and action), and explain their behavior.
You can find more about my work on the Research, Teaching, and CV pages.
Research Interests
Latest News
- Jul 2026 “Breaking the Gaol: What Predicts Jailbreak Success Across Large Language Models” has been accepted to KONVENS 2026 (with Melissa Niemeier and Carolin Holtermann).
- Jun 2026
3 papers accepted:
- Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions (accepted to KI 2026)
- Extraction and Analysis of Multimodal Concepts in Vision Language Models through Sparse Autoencoders (accepted to ICANN 2026)
- Learning to Hear Hesitation: Continual Learning for Disfluency-Aware ASR (accepted to Interspeech 2026)
- May 2026 I have been recognized as a Gold Reviewer for ICML 2026, placing me among the top reviewers based on area-chair ratings.
- Apr 2026 Our paper The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level has been accepted to ICML 2026.
- Apr 2026 Our paper Explaining, Verifying, and Aligning Semantic Hierarchies in Vision-Language Model Embeddings has been accepted to IJCAI 2026.
- Apr 2026 This summer semester, I’ll teach advanced topics in deep learning, which include transformers, large language models, robust generalization, and interpretability (course website).
- Mar 2026
Our DFG grant proposal “Evaluating, Explaining, and Enabling Ethical Multi-Agent Systems of Large Language Models” for the Priority Area LaSTing (SPP 2556) has been approved (with Prof. Anne Lauscher). See our position paper.
We are hiring! - Mar 2026 I am co-organizing the 5th International Workshop on Spatio-Temporal Reasoning and Learning (STRL 2026), which is co-located with IJCAI 2026 and taking place in Bremen, Germany this year.
- Mar 2026 The LUMO project has now an official website with detailed information about the project, the team, and the research. Check it out!
Recent Publications
Learning to Hear Hesitation: Continual Learning for Disfluency-Aware ASR
Annual Conference of the International Speech Communication Association 2026
Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions
German Conference on Artificial Intelligence 2026
Recent Blog Posts
Deep Learning Courses and Books for Beginners
A curated list of deep learning resources for CS students who want to understand recent research papers.
Common LaTeX Mistakes by Newbies
A list of common LaTeX mistakes that beginners often make, with examples of correct usage.