Student researcher · TU Berlin
Interpretability · multilingual NLP · counterfactual explanations
I am a master's student in computer science at TU Berlin, supervised by Qianli Wang. I wrote my bachelor's thesis with Qianli Wang and Nils Feldhus.
My recent research focuses on the interpretability of multilingual language models: whether a concept is represented consistently across languages, and what a model actually relies on when it makes a prediction. Much of my work uses counterfactual explanations, minimal edits to an input that flip a model's prediction. My recent papers improve self-generated counterfactuals through preference optimization in the multilingual setting and through iterative feedback at inference time.
Looking for internship and PhD positions starting in 2027. Feel free to reach out by email.
Two papers were accepted to Findings of EMNLP 2026.
Completed my contribution to Apache SystemDS.
One paper on multilingual counterfactual explanations is in submission.
One paper on counterfactual generation is in submission.
Multilingual NLP
Whether a concept lives in the same place inside a model, no matter which language asks for it.
Interpretability
Pulling a dense representation apart into sparse, individually legible features, most often with a sparse autoencoder.
Counterfactuals
The smallest edit that flips a prediction, and everything that edit gives away.
On-policy distillation
Teaching a small model on the trajectories it actually walks, rather than the ones its teacher would have.
Reinforcement learning
A reward is a rough map. Someone still has to walk the route.
Memory
What a model should carry between turns, and what it is better off forgetting.
Modality priors
Every modality arrives already holding assumptions about what it is looking at.
Findings of EMNLP 2026
From the abstract
Self-generated counterfactual explanations (SCEs) are minimally modified inputs (minimality) generated by large language models (LLMs) that flip their own predictions (validity) […]
Findings of EMNLP 2026
From the abstract
Counterfactual examples are minimal edits to an input that alter a model's prediction. They are widely employed in explainable AI to probe model behavior and in natural language processing (NLP) to augment training data. […]
TU Berlin · course report
From the abstract
We study how much labelled data self-supervised pretraining can save on STL-10, comparing a classical HOG baseline against frozen ViT features from DINOv3 under an identical linear SVM evaluation across varying training-data fractions. […]
TU Berlin · course report
From the introduction
Financial sentiment analysis differs from general opinion mining in a fundamental way: sentiment is rarely conveyed through explicit evaluative words […]
Evidence that I occasionally leave the building.
I contributed to LLM-assisted generation of data augmentation pipelines, building a harness that compared published augmentation baselines against LLM-generated strategies across datasets of different modalities.