Tobias Braun - Zuse School ELIZA

PhD Student
Foundations of ML: Computer Vision

I am a PhD student in Multimodal Artificial Intelligence at TU Darmstadt under the supervision of Marcus and Anna Rohrbach and an ELLIS PhD candidate. My research focuses on trustworthy AI, with particular interests in adversarial robustness, AI security, multimodal learning, and vulnerabilities of unified and agent-based systems. I study how modern AI models can be manipulated and how such weaknesses can be detected and mitigated before deployment. Alongside my research, I lead teaching activities for the Statistical Machine Learning course at TU Darmstadt, one of the university’s largest machine learning courses with more than 500 students, and supervise student research projects. Before my PhD, I worked as an independent ML engineer and conducted more than 100 AI workshops across academia and industry. My most influential recent work is GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows (Grebe, Braun, Rohrbach & Rohrbach, ICML 2026 Spotlight), which introduces a novel geometric framework for robust concept erasure in generative models. I was also recognized as an ICML 2026 Gold Reviewer for outstanding reviewing contributions.

I am a PhD student in Multimodal Artificial Intelligence at TU Darmstadt under the supervision of Marcus and Anna Rohrbach and an ELLIS PhD candidate. My research focuses on trustworthy AI, with particular interests in adversarial robustness, AI security, multimodal learning, and vulnerabilities of unified and agent-based systems. I study how modern AI models can be manipulated and how such weaknesses can be detected and mitigated before deployment. Alongside my research, I lead teaching activities for the Statistical Machine Learning course at TU Darmstadt, one of the university’s largest machine learning courses with more than 500 students, and supervise student research projects. Before my PhD, I worked as an independent ML engineer and conducted more than 100 AI workshops across academia and industry. My most influential recent work is GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows (Grebe, Braun, Rohrbach & Rohrbach, ICML 2026 Spotlight), which introduces a novel geometric framework for robust concept erasure in generative models. I was also recognized as an ICML 2026 Gold Reviewer for outstanding reviewing contributions.

Google Scholar