Markus Peschl

I am a senior deep learning researcher at Qualcomm AI Research in Amsterdam.

I am dedicated to advancing AI research that is both theoretically rigorous and practically impactful. Previously, I worked on reinforcement learning and at the intersection of deep learning and combinatorial optimization for wireless and chip design problems. Lately, my research focuses on generative models for sequence modeling, decision making, and embodied AI.

In my free time, you can find me playing the piano🎹, practicing meditation🧘, reading philosophy📙 and thinking about artificial general intelligence (preferably on the beach🏖️).

Email  /  Scholar  /  LinkedIn  /  X

Research

We use VLMs and diffusion policies with API subroutines as structured supervision to improve generalization and long-horizon robotic control.

Hybrid Training (HyT) lets Vision-Language-Action models learn from chain-of-thought reasoning to boost performance while allowing faster, thought-free inference when needed.

Introducing Oat-VLA: Cut visual tokens for VLAs and train twice as fast as OpenVLA while achieving better real-world performance.

A Geometric Algebra Transformer "WiGATr" for learning an E(3) equivariant simulator of wireless signal transmission. We test WiGATr as a predictive model as well as a diffusion model of signal and 3D geometry.

Using graph transformers to predict Steiner points in a hybrid fashion to tackle the Steiner Tree problem in physical design (chip design).

We use GFlowNets to tackle the NP-hard problem of computation graph scheduling by combining a temparature conditioned policy with top-k sampling.

We combine lattice reduction, Bayesian optimization and a stochastic sampler to arrive at a highly efficient, yet accurate linear MIMO demapper.

AI alignment is challenging due to a multitude of different and possibly conflicting values present in human feedback. We propose to overcome this challenge with a multi-objective agent that actively learns user preferences.