Darius Muglich

I’m a DPhil student at Oxford working on machine learning and optimization.

Selected works

Full list on Scholar
  1. 2026

    Ad-Hoc Teamwork from Human Demonstrations

    Darius Muglich, Niklas Lauffer, Tin Dizdarevic, Jakob Foerster

    NeurIPS 2026 PDF and code coming soon
  2. 2025

    Expected Return Symmetries

    Darius Muglich, Johannes Forkel, Elise van der Pol, Jakob Foerster

    ICLR 2025
    PDF Code
  3. 2022

    Equivariant Networks for Zero-Shot Coordination

    Darius Muglich, Christian Schroeder de Witt, Elise van der Pol, Shimon Whiteson, Jakob Foerster

    NeurIPS 2022
    PDF Code
  4. 2022

    Generalized Beliefs for Cooperative AI

    Darius Muglich, Luisa Zintgraf, Christian Schroeder de Witt, Shimon Whiteson, Jakob Foerster

    ICML 2022
    PDF Code

Current Work

Neurosymbolic robot control
Robots typically run hand-engineered controllers that struggle in situations their designers didn’t anticipate. I’m developing methods that refine existing controllers, such as behaviour trees, with learned adaptations. The robot improves from experience while keeping an explicit decision structure that engineers can inspect and formally verify.
Logic gate network optimization
Logic gate networks are compact models that run very efficiently on hardware, but improving a trained circuit is hard because useful changes often couple a connection and a gate function. I’m developing a discrete optimization method that makes these joint edits practical to evaluate, with efficient approximations for when exhaustive scoring is expensive.
Implicit bias in reinforcement learning
When many policies are equally good under the training objective, the optimizer’s implicit preferences determine which one is learned, and with it how the policy generalizes. I’m developing theory that characterizes this implicit bias for policy-gradient methods.