Launching GRASP, a new AI research chair on interpretability of multimodal LLMs
Published:
I am very happy to announce the start of GRASP (Graph Signal Processing for Interpreting and Steering Multimodal Large Language Models), a research chair funded by the Cluster SequoIA as part of France 2030, which I hold at IMT Atlantique within the BRAIn team of Lab-STICC.
The problem. Multimodal LLMs that take audio, images or brain signals as input sometimes ignore them: they rely on their internal linguistic priors rather than on the sensory evidence they are given. We call this failure mode perceptual decoupling, and it is a real obstacle to using these models in high-stakes domains such as health, neurotechnology or environmental monitoring.
The idea. GRASP treats a neural architecture as a communication graph and its activations as signals on that graph. Graph Signal Processing then lets us decompose internal computation into interpretable spectral components, and spectral filters become a lever for intervention. This transposes to AI models a notion I have long worked with in neuroimaging: structure-function coupling, i.e. how the brain’s anatomical connectivity constrains its activity.
« Neuroimaging research taught us to measure and interpret the gap between a structure and the activity flowing through it. I propose to apply the same lens to AI models. »
The chair has three objectives:
- A GSP framework for mechanistic interpretability: model Transformers and state-space models as graphs, and quantify the coupling between their topology and their activation dynamics, going beyond atomised features such as those found by sparse autoencoders.
- Diagnosing and steering multimodal integration: detect when a model’s reasoning decouples from its sensory inputs, and introduce filter-based steering to re-couple it to perception, without retraining or fine-tuning.
- Validation on ecoacoustics and neuroimaging: audit audio-language models for bioacoustics (Silent Cities, Observatoire de l’Environnement Nocturne, PEPR Dynabiod) and models combining video, text, audio and brain activity (Courtois NeuroMod).
The team. GRASP brings together Bastien Pasdeloup and Axel Marmoret (IMT Atlantique), Vincent Miele (CNRS, LECA) and Jérémy Froidevaux (Université de Franche-Comté), postdoc Venkatesh Subramani, PhD students Nolan Sisouphanthong and Kaiteng Jiang, and collaborators Giulia Lioi (IMT Atlantique), Lune Bellec and Karim Jerbi (Université de Montréal, MILA).
We are recruiting! Several positions at the undergradutate, graduate, and postdoc positions are currently open. Feel free to get in touch if you are interested.
This work is funded by the « Agence Nationale de la Recherche » through the France 2030 programme under the reference number ANR-23-IACL-0009.
