Recent news

Marvin Lavechin

Posted on by

Marvin Lavechin has recently joined CNRS as chargé de recherche at the Laboratoire d’Informatique et Systèmes (LIS) in Marseille. He completed his PhD at Meta AI and the Laboratoire de Sciences Cognitives et Psycholinguistique (ENS, Paris), followed by postdocs at GIPSA-lab (Grenoble), and then at the Computational Psycholinguistics Lab (MIT) and the Bergelson Lab (Harvard), funded by the Simons Center for the Social Brain.

His research sits at the intersection of artificial intelligence and cognitive science. Marvin seeks to understand how children learn to speak and perceive language, and to reproduce this process in machines. He develops automatic tools to analyze what children hear and vocally produce in their daily lives (see, for example, BabAR). Marvin also designs computational models that “learn” language the way an infant would, from raw sounds, without predefined labels or categories, in order to better identify the mechanisms that make this acquisition possible.

This research has both fundamental and practical implications: better understanding language acquisition can improve early screening for developmental disorders, while also opening new avenues for designing more efficient artificial intelligence.

Clarification-request feedback provides a learning signal for grammar development

Posted on by

In natural child–caregiver conversations, caregivers are more likely to ask for clarification after a child says something ungrammatical, such as “I goed,” than after a grammatical utterance, such as “I went”, shown in Panel A. This means that clarification requests carry information about whether the child’s sentence was well formed. In real conversations, input and feedback to children are tightly correlated, so it is hard to know whether feedback itself adds anything beyond the language children already hear.

We leveraged computational modelling (GPT-2) to test whether clarification requests can actually help learning. By training GPT-2 on the same linguistic input that children hear, either with or without the clarification-request feedback they receive in conversation, we isolated the specific contribution of feedback to grammatical learning. Panel B shows that the model trained with feedback develops better grammatical language, suggesting that everyday conversational responses in children’s experience can provide a useful learning signal.

Mitja Nikolaus and Abdellah Fourtassi. 2026.
Philosophical Transactions B 381 (1943): 20240374  —  @HAL

Prof. Sonja Kotz

Posted on by
Prof. Sonja Kotz

Prof. Sonja Kotz — photo (c) IMERA

Prof. Sonja Kotz, a long-standing member of the ILCB International Advisory Board, is starting a sabbatical at Aix-Marseille Université this month.

Sonja Kotz is Professor of Neuropsychology and Translational Cognitive Neuroscience at Maastricht University (The Netherlands). Her research explores body-brain-behaviour dynamics related to temporal (when) and content (what) predictions in audition, speech, and music across the lifespan, as well as in clinical populations including Parkinson’s disease, stroke, psychosis, tinnitus, and dyslexia. In this research she uses a broad range of behavioural and neuroimaging methods (M/EEG, r/s/fMRI, TMS). She previously served as President of both the European Society for Cognitive and Affective Neuroscience (ESCAN) and the Society for the Neurobiology of Language (SNL) and is a senior editor at several domain specific journals (Imaging Neuroscience, Neurobiology of Language).

Sonja Kotz collaborates extensively with many leading researchers in speech, music, and cognitive neuroscience at the ILCB and worldwide.