Marcin Sendera

I am a Research Assistant at Jagiellonian University (in the GMUM group, advised during my PhD by Prof. Jacek Tabor) and an AI Research Scientist at NASK – National Research Institute, focusing on AI Safety. Previously, I was a Research Intern at Mila - Quebec AI Institute (2023-2024), under the supervision of Prof. Yoshua Bengio.

In May 2026 I successfully defended my PhD thesis, “Probabilistic deep learning: from efficient sampling to principled generation,” which was awarded with distinction by the committee. The thesis was reviewed by Prof. Arnaud Doucet (University of Oxford / Google DeepMind), Prof. Amos Storkey (University of Edinburgh), and Prof. Piotr Miłoś (University of Warsaw / IMPAN / Mistral AI). The full thesis is available here.

My research focuses on Probabilistic Deep Learning, with a specific emphasis on Diffusion Models and advanced sampling methods like GFlowNets. My core mission is to leverage these tools to build AI models that possess stronger Probabilistic Reasoning capabilities while remaining Safe and aligned. I am particularly interested in how structured generation can bridge the gap between rigorous Bayesian inference and modern deep learning.

I am also trying to advance the (amortised) variational inference methods to allow better probabilistic inference (and Bayesian posterior estimation) in complex, large-scale models like LLMs and Diffusion Models.

Current Research Focus:

Reasoning: Improving reasoning capabilities in generative models via structured sampling.

Diffusion: Developing differentiable Top-K methods for discrete data generation and improving continuous diffusion models.

Safety & Unlearning: Embedding safety constraints and machine unlearning directly into the generative process.

Prior to my PhD, I was a Summer Research Student at the University of Cambridge, working on deep learning for early cancer detection from DNA methylation data.