I currently serve as a technology specialist focusing on AI Safety at the AI Office in the European Commission, where I help shape and enforce policies to ensure the responsible development of artificial intelligence systems.
Previously, I was a PhD candidate at the Institute for Logic, Language and Computation (ILLC) at the University of Amsterdam. My doctoral research investigated the mechanisms behind social biases in language models—examining how these biases manifest, how they can be measured reliably, and grounding these technical discussions in broader societal contexts.
I'm particularly passionate about leveraging interpretability tools to address critical questions like "How can we reliably measure and mitigate bias?" and "How do language models acquire these biases during training?" Through this work, I aim to make AI systems more equitable, transparent, and aligned with human values.
During my doctoral studies, I contributed to research on bias and interpretability through collaborations with EleutherAI and the BigScience initiative to advance the field of responsible AI.
AI Safety · Bias in NLP · Interpretability
∗ indicates equal contribution
Bloom: A 176b-parameter open-access multilingual language model
arXiv, abs/2211.05100
The birth of bias: A case study on the evolution of gender bias in an english language model
4th Workshop on Gender Bias in Natural Language Processing (GeBNLP), p. 75
Undesirable biases in NLP: addressing challenges of measurement
Journal of Artificial Intelligence Research
Pythia: A suite for analyzing large language models across training and scaling
ICML 2023, vol. 202, pp. 2397–2430. PMLR
BlackboxNLP 2023, pp. 379–394
Chapgtp, illc's attempt at raising a babylm: Improving data efficiency by automatic task formation
BabyLM Challenge at CoNLL 2023, pp. 74–85
Inseq: An interpretability toolkit for sequence generation models
ACL 2023 (System Demonstrations), pp. 421–435, Toronto
You reap what you sow: On the challenges of bias evaluation under multilingual settings
BigScience Workshop at ACL 2022, pp. 26–41
The grammar of emergent languages
EMNLP 2020, pp. 3339–3359
This post is about our paper "Undesirable Biases in NLP: Addressing Challenges of Measurement", published in JAIR. Developing tools for measuring & mitigating bias is challenging: LM bias is a complex sociocultural phenomenon and we have no access to a ground truth. We voice our concerns about current bias evaluation practices…
In my research, I use various approaches to investigate social bias in language models. When discussing such undesirable biases, we often take a mathematical and 'mechanistic' approach, measuring deviations from a prescriptive norm of ideal behavior (e.g., a skew in gender distribution from 50/50%) or trying to explain…
Language models have become essential building blocks of modern AI systems dealing with natural language. These models excel in diverse tasks including sentiment analysis, text generation, translation, and summarization. While their effectiveness stems from neural network architectures trained on vast datasets, this power comes with significant challenges…