For a complete and up-to-date list of my papers, please see my Google Scholar profile. A selection of my work is listed below.
Published in ICML, 2026
We study practical in-training safeguards against emergent misalignment, evaluating regularization, safe subspaces, and data interleaving.
Recommended citation: David Kaczér, Magnus Jørgenvåg, Clemens Vetter, Esha Afzal, Robin Haselhorst, Lucie Flek, Florian Mai. (2026). "In-Training Defenses against Emergent Misalignment in Language Models." ICML 2026. https://arxiv.org/abs/2508.06249
Published in IASEAI'26: International Association for Safe and Ethical AI Conference, 2026
We analyze overlap in failure modes across alignment techniques to assess the limits of defense-in-depth risk mitigation.
Recommended citation: Leonard Dung, Florian Mai. (2026). "AI Alignment Strategies from a Risk Perspective: Independent Safety Mechanisms or Shared Failures?" IASEAI'26: International Association for Safe and Ethical AI Conference. https://arxiv.org/abs/2510.11235
Published in COLM, 2024
We propose a method to learn planning for language modeling using unlabeled data.
Recommended citation: Nathan Cornille, Marie-Francine Moens, Florian Mai. (2024). "Learning to Plan for Language Modeling from Unlabeled Data." COLM 2024. https://arxiv.org/abs/2404.00614
Published in ACL, 2023
We propose an efficient all-MLP architecture with the same inductive biases as Transformers.
Recommended citation: Florian Mai, Arnaud Pannatier, Fabio Fehr, Haolin Chen, François Marelli, François Fleuret and James Henderson. (2020). "HyperMixer: An MLP-based Low Cost Alternative to Transformers." ACL 2023. https://arxiv.org/abs/2203.03691
Published in EMNLP, 2020
We reduce conditional text generation tasks to learning in the embedding space of an autoencoder.
Recommended citation: Florian Mai, Nikolaos Pappas, Ivan Montero, Noah A. Smith and James Henderson. (2020). "Plug and Play Autoencoders for Conditional Text Generation." EMNLP 2020. https://arxiv.org/abs/2010.02983
Published in ICML, 2020
We formulate a benchmarking evaluation protocol that takes the tunability of optimizers into account.
Recommended citation: Prabhu Teja Sivaprasad*, Florian Mai*, Thijs Vogels, Martin Jaggi and Francois Fleuret (2020). "Optimizer Benchmarking Needs to Account for Hyperparameter Tuning." ICML 2020. https://arxiv.org/abs/1910.11758
Published in ICLR, 2019
We represent word embeddings as matrices and compose them via matrix multiplication.
Recommended citation: Florian Mai, Lukas Galke and Ansgar Scherp (2019). "CBOW Is Not All You Need: Combining CBOW with the Compositional Matrix Space Model." ICLR 2019. https://arxiv.org/abs/1902.06423