@SebSanokowski

Post-Doc at TU Munich, PhD in Artificial Intelligence. Working on diffusion samplers.

Joined November 2021
  • Tweets91
  • Following146
  • Followers220
12Media
Learning to remove noise from images yields fantastic image generators. Here's how you can use diffusion models to solve Combinatorial Optimization problems by removing noise from solutions that you don't even have! To be presented @ICML 2024, 🧵👇
3
54
5
210
54,584
Sebastian retweeted
Sebastian@SebSanokowskiOct 15
1/14 Are you wondering if you can turn your favourite maximum-entropy RL algorithm into a diffusion-based variant that learns multimodal action distributions? It turns out the answer is yes, using Diffusion-Augmented Markov Decision Processes (DA-MDPs).
1
1
2
69
1/14 Are you wondering if you can turn your favourite maximum-entropy RL algorithm into a diffusion-based variant that learns multimodal action distributions? It turns out the answer is yes, using Diffusion-Augmented Markov Decision Processes (DA-MDPs).
1
1
2
69
more replies
13/14 Importantly, DA-MDPs also work with action chunking, widely used in diffusion-based robot control. We demonstrate policies that learn to generate multiple actions jointly and execute them in sequence.
1
18
14/14 Paper: arxiv.org/abs/2512.02019 JAX: github.com/sanokows/Diffusio… PyTorch: github.com/Atarilab/DA_MDP_p… #ReinforcementLearning #DiffusionModels #Robotics

Diffusion-Augmented Markov Decision Processes for Maximum Entropy...

Diffusion models provide an expressive framework for sampling from complex, unnormalized distributions. In this work, we extend Maximum Entropy Reinforcement Learning (ME-RL) to diffusion-based...

arxiv.org
1
8
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.
3
19
1
120
7,061
Ever experienced instabilities when using the popular LV (Log Variance) loss for training Diffusion Bridge Samplers?
1
6
3
18
10,684
more replies
If you’re at NeurIPS in San Diego next week, we’d love to connect and discuss our work. Joint work with: @lugruber0, Christoph Bartmann, @HochreiterSepp, @sebaLeh
1
1
3
215
Links: neurips.cc/virtual/2025/loc/… arxiv.org/abs/2506.10982
1
3
156
Load more