Cognitive neuroscience and language processing
PhD at NeuroSpin under Prof. Stanislas Dehaene. Flagship output: a Cortex paper on hierarchical versus sequential computations during sentence reading (MEG and EEG, compared to an LSTM language model). Separate first-author work on grammatical agreement (EMNLP 2023) and LLM evaluation papers.
Cortex (2026)
First author on Disentangling Hierarchical and Sequential Computations during Sentence Processing (Cortex, Elsevier), with Stanislas Dehaene and Yair Lakretz. Twenty-two participants read sentences in a rapid serial visual presentation (RSVP) task while we recorded combined magnetoencephalography (MEG) and electroencephalography (EEG) at NeuroSpin and ICM, Paris.
Multivariate decoding of MEG and EEG: only hierarchical (structural) effects were above chance; local transition and agreement congruity stayed at chance. A two-layer LSTM on the same sentences showed decodable structure, transition, and congruity effects. The paper concludes that human sentence processing is dominated by structure-based computation, unlike this recurrent language model. Figure on the publications page.
PhD, NeuroSpin / Sorbonne University
Supervised by Prof. Stanislas Dehaene (Collège de France). Designed and ran neuroimaging experiments with EEG and MEG; additional intracranial work with collaborators is outside the Cortex paper above.
EMNLP 2023
First author on a main-track paper at EMNLP 2023 (top 14% of submissions): grammatical agreement in humans and language models. Co-authored with researchers from Meta AI and NeuroSpin. Completed alongside clinical machine learning work on intravascular biosignals.
First-author language-model evaluation work at IJCNLP 2025 (personality traits in large language models) and Springer AICS 2025 (semantic violation detection in causal LMs). Figures and detail on the publications page.