About me
Short Biography
Florian is a Postdoctoral Research Fellow at the Language Intelligence & Information Retrieval lab at KU Leuven, led by Marie-Francine Moens, where he works on learning planning components for language modeling from unlabeled data.
He received his PhD from EPFL and worked in the Natural Language Understanding group at Idiap Research Institute led by James Henderson. His PhD thesis aims to decrease the costs of natural language understanding in terms of processing time (CMOW sentence embeddings and HyperMixer), required labeled data, and hyperparameter tuning.
During his master’s studies, he worked as a student research assistant with Ansgar Scherp at Leibniz Information Centre for Economics - ZBW. He studied several NLP-related topics like topical document classification in digital libraries and citation recommendation.
Current Research Interest
Language models have become so successful because we can train them effectively from large amounts of diverse, unlabeled data, which enables them to acquire a wide variety of skills and knowledge. This is a reflection of the ‘bitter lesson’: Any algorithmic approach to AI needs to be able to easily scale to large amounts of (unlabeled) data.
Planning and reasoning remains a core challenge for LLMs. However, despite the teachings of the bitter lesson, the vast majority of current approaches to planning and reasoning in LLMs do not try make use of the unlabeled data. In my opinion, taking the self-supervised learning approach to the level of planning and reasoning is the key to unlocking the full potential of LLMs.
During my time as a postdoc at KU Leuven, I led a project to learn planning components for language modeling from unlabeled data. In our first published work, we learn a planning component to predict an abstract writing action of the next sentence, which guides the language model in its predictions of the future writing process. In the future, we plan to extend this approach to predicting sequences of writing actions, and to integrate RL-based planning towards a writing goal.
Personal
When I am not obsessing about AGI on Reddit pondering about the future of artificial intelligence, I binge Netflix with my gf spend time with my family, torture my overweight body on foot, on bike and in water practice for a triathlon, or waste countless hours in Counter Strike 2 hone my strategic thinking skills and hand-eye coordination in virtual simulations.