Basel Mousi

MSc Data Science student at EPFL. NLP researcher working on multilingual and multimodal LLMs, interpretability, and knowledge editing.

I'm a Master's student in Data Science at EPFL. Before that, I spent nearly four years as a Research Engineer at the Qatar Computing Research Institute (QCRI), where I led projects on multilingual knowledge editing, multilingual reasoning, simultaneous speech translation, and benchmarks for Arabic dialects and culture.

My research asks how large language models represent and use knowledge across languages and modalities, and where that breaks down. I've worked on interpreting the internals of language models, on editing what models know so the change carries across languages, and on hallucination and instability in vision-language and speech models. Much of this work focuses on Arabic and its dialects. My papers have appeared at ACL, EMNLP, COLING, EACL and Interspeech.

I hold a BE in Electrical and Computer Engineering from the American University of Beirut, where I graduated with distinction.

Research interests

Multilingual LLMs and cross-lingual generalization
Interpretability
Knowledge editing
Hallucination and robustness in multimodal (vision and speech) models
Arabic NLP

News

All news

Sep 2026Started my MSc in Data Science at EPFL.
Aug 2026Wrapped up nearly four years as a Research Engineer at QCRI.
May 2026"Said Aloud, Read Different" accepted at Interspeech 2026.
May 2026"Once Correct, Still Wrong", on counterfactual hallucination in multilingual VLMs, accepted to Findings of ACL 2026. The M2CQA benchmark is on Hugging Face.
Mar 2026Fanar 2.0 technical report released.

Featured projects

All projects

Reliability of Multilingual Multimodal Models
Where vision- and speech-language models fail across languages, modalities and cultures.
Team Lead · QCRI
Shaheen: Machine Translation for Arabic
English↔Arabic and dialectal machine translation models used to generate training data for Fanar.
Developer · QCRI
LLMeBench
An open-source framework for benchmarking LLMs on any NLP task, in any language.
Core Contributor · QCRI
NeuroX: Interpretability Toolkit
A toolkit for analyzing and interpreting what neurons and latent concepts in language models encode.
Core Team Member · QCRI
AraDiCE: Arabic Dialect and Cultural Evaluation
Benchmarks testing how well LLMs handle Arabic dialects and regional cultural knowledge.
Team Lead · QCRI