Builder and ML Research Engineer at the ELLIS Institute Tübingen, working on post-training of multilingual LLMs. Previously, I was a PhD student at Frank Hutter's Machine Learning Group in Freiburg.
Current projects / research interests:
- Can LLMs Beat Classical Hyperparameter Optimization Algorithms? A Study on autoresearch (1st author) [Code] [Interactive demo]
- OLMo 3 post-training analysis: how does think/instruct capability evolve over sft training time? See this report
- What dataset mixtures allow LLMs to improve multilingual support (without curbing English language support)? See this report
Past projects:
- Improving LLM-based Global Optimization with Search Space Partitioning (4th author, ICLR 2026)
- Beyond Random Augmentations: Pretraining with Hard Views (1st author, ICLR 2025) [code]
- Quick-Tune: Quickly Learning Which Pretrained Model to Finetune and How (2nd author, ICLR 2024, oral) [code]
- Zero-Shot AutoML with Pretrained Models (1st author, ICML 2022, spotlight) [code]
- Learning Environments for Reinforcement Learning (1st author, ICLR 2022) [code]
My PhD thesis: Meta-Learning and Synthetic Data for Automated Pretraining and Finetuning (supervised by Prof. Frank Hutter)
- allenai/open-instruct
- Post-training recipes for language model (see this PR)
- OpenEuroLLM/llm-judge-eval – LLM-as-judge evaluation framework
🔬 autoresearch-automl: Can LLMs Beat Classical Hyperparameter Optimization Algorithms? A Study on autoresearch [Paper]
🔍 aretheyinvolved.com: Search + analyse names across all Epstein files (NER, OCR, co-occurrences, AI summaries+role classification).
🌸 myperfumeai.com: ChatGPT for perfumes / personal perfume recommendation AI assistant based on 120k+ perfumes dataset.
🧠 Conditional Density Estimation (CDE) – Reference package (cde) for conditional density estimation.
🧩 slurmfrag: Fine-grained experiment fragmentation and management for SLURM clusters.
📊 tailgrid: Terminal-based and LLM-assisted monitoring and visualization of log files.
🧮 Generative Symbolic Regression: Neural translation from tabular data to concise LaTeX equations. (private / research code)
🤖 mppi_pendulum: Minimal MPPI control implementation for the classic pendulum task.
🎯 Zero-Shot AutoML with Pretrained Models: Zero-shot selection of strong pretrained models without training.
🎥 video2tfrecord: Convert raw video datasets into scalable TFRecord pipelines.