Denis Blessing

🌻 One final paper for 2025 :) really excited about this work with @PPotaptchik and @brianlee_lck. We show that there is an analytical ODE connecting a flow/map/diffusion model to its tilted alternative. Allows us to turn fine-tuning into an iterative regression! Works well and has nice scaling properties :) check it out, and stay tuned for more!

Best paper award "Learning Boltzmann Generators via Constrained Mass Transport" @DenBless94 and Christopher von Klitzing openreview.net/forum?id=1mG9h…

Presenting our spotlight paper on trust regions for optimal control at NeurIPS, arxiv.org/pdf/2508.12511. We show that KL-equipspaced measure transport can be interpreted as geometric annealing with adaptive step sizes, leading to major performance gains on hard control problems.

If you're at NeurIPS this week and are interested in stochastic optimal control (SOC) and diffusion models come by our poster on Friday, Dec 6 • 4:30 - 7:30 PM Joint work with, @julberner, @cdomingoenrich, @YuanqiD, @ArashVahdat and Gerhard Neumann.

When sampling from multimodal distributions, we rely on multiple temperatures to balance exploration and exploitation. Can we bring this idea into the world of diffusion-based neural samplers? 👉Check out our ICML paper to see how this idea can lead to significant improvements!

Excited for #ICML2025 in Vancouver! On Thursday morning, I'm presenting our paper (arxiv.org/pdf/2506.00962) on a critical issue in reinforcement learning: how to correctly handle random time horizons. We've identified incorrect formulas and offer a solution. Let's chat, write me!

1/ Where do Probabilistic Models, Sampling, Deep Learning, and Natural Sciences meet? 🤔 The workshop we’re organizing at #NeurIPS2025! 📢 FPI@NeurIPS 2025: Frontiers in Probabilistic Inference – Learning meets Sampling Learn more and submit → fpiworkshop.org Organizers: @tara_aksa @msalbergo @bose_joey @marylougab @theh2o64 @guanhorng_liu Grant Rotskoff @EvaSmorodina @AlexanderTong7

I am happy to share that our work ‘DIME: Diffusion-Based Maximum Entropy Reinforcement Learning’ has been accepted to ICML 2025. Many thanks to my colleagues and collaborators @softraeh, @DenBless94, @BruceGeLi, @DPalenicek, @Jan_R_Peters, @GeorgiaChal,@geri_neumann

Our new work arxiv.org/pdf/2503.01006 extends the theory of diffusion bridges to degenerate noise settings, including underdamped Langevin dynamics (with @DenBless94, @julberner). This enables more efficient diffusion-based sampling with substantially fewer discretization steps.