Yusuf Malik Durmus

Preference-Guided Curriculum Discovery for Robust RL

Bachelor's thesis · LMU Munich, Chair of Mobile and Distributed Systems · 2026

Reinforcement learning agents generalize better when they train on well-chosen environments. I used an LLM (Gemini) to compare about 88,000 pairs of rendered levels by difficulty, validated the labels against human judgment, and trained a neural network ensemble that ranks new levels.

I then extended a state-of-the-art curriculum algorithm (Sampling for Learnability) with this difficulty model, new selection rules and an adaptive difficulty target, and evaluated it in about 200 GPU-accelerated training runs with paired statistical tests.

• JAX / Flax
• Reinforcement learning
• Preference learning
• Gemini on Vertex AI
• Slurm

• Project report
• Thesis (PDF)
• Code

jax-keyguard: Catching PRNG Key Reuse in JAX

Open-source Python package · Personal project · 2026

In JAX, reusing a random key silently repeats the same "random" numbers. I built a tool that finds this mistake in two ways: a static checker that follows keys through branches, loops and aliases without running the code, and a runtime check that raises an error the moment a key is reused.

It covers the older PRNGKey keys that JAX's own check skips. Run on the JAX repository, it found an example program that draws every layer's weights and biases from the same key.

• Python
• JAX
• Static analysis
• pytest / flake8 plugins
• PyPI

• Project report
• Code
• PyPI

rng = jax.random.PRNGKey(0)
w = jax.random.normal(rng, (m, n))
b = jax.random.normal(rng, (n,))

KG001 PRNG key 'rng' reused:
already consumed by
jax.random.normal on line 173

RLHF-Gym: Reinforcement Learning from Human Feedback

University software engineering project · LMU Munich · 2023–2024 · Grade 1.0

A complete pipeline that teaches robotic agents in MuJoCo from human preferences instead of a hand-written reward. I led the algorithm work: implementations of PEBBLE and SURF on top of Soft Actor-Critic, a learned reward model with semi-supervised learning and data augmentation, and ensemble-based query selection.

The project includes a web interface where people compare two agent videos, a thread-safe backend that feeds their answers into training, and training jobs on the university Slurm cluster.

• PyTorch
• SAC
• Flask
• MuJoCo
• Slurm

• Code

Inferno Swarm: Autonomous Multi-Agent Simulation

Personal project · 2025

A C++17 simulation of autonomous agents with emergent swarm behaviour. I implemented the steering behaviours (separation, alignment, cohesion) and force accumulation from scratch, and a cellular-automaton environment in which fire spreads and agents switch between flocking and extinguishing.

• C++17
• SFML
• CMake / vcpkg