معرفی
Jonas Geiping is a Professor at the ELLIS Institute Tübingen, leading the Safety- and Efficiency- aligned Learning research group. He has previously worked at the University of Maryland, University of Siegen, and University of Münster.
His research focuses on safety and efficiency in machine learning, exploring data poisoning principles, watermarking for generative models, privacy in federated learning, and adversarial attacks against large language models. He also investigates how to make AI systems more efficient through weight averaging, recursive computation, and computational constraint analysis.
His recent work includes test-time computation scaling using recurrent depth approaches, adversarial attacks on LLMs beyond jailbreaking, and zero-shot detection of machine-generated text through contrasting LLMs (Binoculars method). He has contributed to arithmetic capabilities in transformers with positional embeddings and scalable LLM training frameworks like AxoNN.
His research addresses critical questions about the intersection of safety and efficiency in AI systems: Can models reason well without sacrificing safety? How do computational constraints affect safety guarantees? Can systems be designed where intelligence and safety reinforce each other?
