Research
Welcome! I am a PhD student at the German Research Center for Artificial Intelligence (DFKI) and Saarland University. I work in the E&E group (Efficient and Explainable NLP Models, led by Simon Ostermann) within DFKI's Multilinguality and Language Technology department, contributing to the TRAILS and SOOFI projects. I am interested in how large language models represent and use knowledge across languages and cultures, and in the tools that let us look inside them.
My current work centers on three threads:
- Mechanistic interpretability for multilingual models. I train and analyze sparse autoencoders (SAEs) across languages and architectures, from Llama, Qwen, and Gemma to hybrid Mamba-attention models, to understand which features are shared across languages and which are language-specific.
- Steering and controlled generation. Building on interpretability, I develop activation-steering methods that push multilingual models toward cultural knowledge, better low-resource synthetic data, and controlled linguistic behavior such as tense.
- Culturally sensitive and inclusive NLP. I study how cultural and locale-specific knowledge is encoded in models, including contamination-robust translation benchmarks, stereotype tracing, and dialectal Arabic speech processing.
Earlier in my PhD I worked on document-level neural machine translation, exploring how web-crawled parallel data such as ParaCrawl can be used beyond the sentence level.
Multilingual NLP · interpretability · sparse autoencoders · activation steering · machine translation · cultural knowledge in LLMs
Publications
Also on Google Scholar.
2026
- Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection 6th Workshop on Trustworthy NLP (TrustNLP 2026)
- DFKI-MLT at SemEval-2026 Task 7: Steering Multilingual Models towards Cultural Knowledge 20th International Workshop on Semantic Evaluation (SemEval 2026)
- CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmark arXiv:2601.08331
- Tracing Stereotypes from Representation to Output in Multilingual LLMs arXiv:2609.08322
- Grounding or Guessing? Visual Signals for Detecting Hallucinations in Sign Language Translation International Conference on Learning Representations (ICLR 2026)
- Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation arXiv:2606.18389
- Cultivar: A Contrastive and Locale-Oriented Translation Benchmark for Investigating Contamination and Localisation Robustness arXiv:2608.09766
2025
- TenseLoC: Tense Localization and Control in a Multilingual LLM 5th Workshop on Multilingual Representation Learning (MRL 2025)
- Saarland-Groningen at NADI 2025 Shared Task: Effective Dialectal Arabic Speech Processing under Data Constraints Third Arabic Natural Language Processing Conference (ArabicNLP 2025)
- Language Arithmetics: Towards Systematic Language Neuron Identification and Manipulation 14th International Joint Conference on Natural Language Processing (IJCNLP-AACL 2025)
- Modular Arithmetic: Language Models Solve Math Digit by Digit 14th International Joint Conference on Natural Language Processing (IJCNLP-AACL 2025)
2023
- Exploring Paracrawl for Document-level Neural Machine Translation 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2023)
Software & Models
I release most of my research artifacts openly, with over 100 models and 20 datasets on Hugging Face, and code on GitHub. Highlights:
- nemotron-lens TransformerLens-style interpretability tooling for NVIDIA Nemotron hybrid models (Mamba-2 + attention + MoE): hook any location, cache activations, train JumpReLU SAEs, and evaluate them with SAE-Bench (PyPI).
- MULTI21-SAES Multilingual sparse autoencoder suites for Qwen3, Llama 3.1, and Gemma 2, from the work on multilingual steering by design.
- FineWeb-CLaR Culture- and region-annotated corpora derived from FineWeb, for studying cultural and locale signals in pretraining data.
- Cultural NLP Hub A living, filterable catalog of cultural-NLP datasets with multi-annotator coding and a community submission workflow (code).
Service
- Member of the E&E group (Efficient and Explainable NLP Models) at DFKI's Multilinguality and Language Technology department.
- Researcher on the TRAILS project (Trustworthy and Inclusive Machines), funded by the German Federal Ministry of Research, Technology and Space.
- Researcher on the SOOFI project, building European foundation models on European infrastructure (IPCEI-CIS / 8ra). My work includes the Soofi-SAEs interpretability suite.
- Shared-task participant and system builder: SemEval-2026 Task 7 (cultural knowledge) and NADI 2025 (dialectal Arabic speech processing).
- Open-source contributor to the multilingual interpretability ecosystem, including SAE training pipelines, benchmarks, and datasets.