Jean-Rémi King

Novel study: Emergence of Phonemic, Syntactic, and Semantic Representations in Neural Networks.
Joint work with P. Diego supervision: Y. Lakretz, E. Chemla, Y. Boubenec, @jeanremiking.bsky.social
We explore the emergence of 3 linguistic structures in Neural Networks.
Arxiv: arxiv.org/abs/2601.18617

📣 Our latest brain-to-text decoding results from our Brain team is out:

"Towards decoding individual words from non-invasive brain recordings"

📄 www.nature.com/articles/s41...

👥 Led by Stéphane d'Ascoli & w/ Corentin Bel, Jérémy RAPIN, Hubert Banville, Yohann Benchetrit and Christophe Pallier

A 10-hour within-participant magnetoencephalography narrative dataset to test models of language comprehension - Scientific Data Measurement(s) brain activity measurement Technology Type(s) Magnetoencephalography Factor Type(s) auditory story Sample Characteristic - Organism Homo sapien

Thanks to Meta and ENS for their support, all Open Sources models at HuggingFace, as well as K Armeni, JM Schoffelen for the great MEG dataset (nature.com/articles/s41...), thanks to the community and see you at
@NeuripsConf !

This study strengthens a surprising phenomenon:
Even though brains and LLMs differ in many ways (architecture, modality, and learning goals), they converge on similar sequence of computations.

Understanding why this convergence emerges may help bridge biological and AI.

Two more facts:
1. This temporal alignment is not explained by word predictability.
2. bidirectional models (not inferential like the brain) do not converge toward similar sequential dynamics.

However, this temporal alignment does not emerge automatically.
For the same model and configuration, it seem to depend on:
• Model size (larger → more brain-like)
• Context length (longer context → more brain-like)
Untrained models show almost no alignment.

LLMs build representations in the same order as the brain:
➡️ First layers ↔ early brain responses
➡️ Deep layers ↔ later brain responses
This temporal alignment is robust (r=0.99, p<1e-6) and holds across architectures (transformers & recurrent), sizes, and training regimes.

🧠Question: How similar are the hierarchical computations between LLMs and the human brain?
Approach: we compare the evolution of MEG brain recordings across time to the evolution of latent representations across layers of 17 LLMs.

🎉 Our paper has been selected for a Neurips Spotlight:

“Scaling and Context Steer LLMs along the Same Computational Path as the Human Brain”

👥led by J Raugel, w/ S. Ascoli, Rapin & @valentinwyart.bsky.social

📄https://openreview.net/pdf?id=4YKlo58RcQ
📍 Hall C-E Poster #2006
🧵thread 👇

🧠How does the hierarchy of speech representations unfolds in the human brain?

Our latest work, led by @lauragwilliams.bsky.social, together with Alec Marantz and @davidpoeppel.bsky.social, is now out in PNAS:

www.pnas.org/doi/10.1073/...

Jean-Rémi King - The Emergence of Language in the Human Brain

We’re excited to announce that @jeanremiking.bsky.social , CRNS researcher at the ENS Paris & leader of Meta’s Brain & AI team, will be sharing groundbreaking insights into how language emerges in the human brain at BHD2025!

[2/2] 📊 His research uses encoding and decoding approaches to show how modern speech and language models account for brain responses to natural speech, measured with EEG, MEG, iEEG, and fMRI, even in children aged 2 to 12.

📆 November 19–21, 2025.

+info 👇

brainhack-donostia.github.io

Congrats Mariya!

Thanks to all the great researchers who contributed to this project: Joséphine Raugel, the DINOv3 team, @valentinwyart.bsky.social, FAIR and ENS as well as the open source and open data #NeuroAI communities for making this possible! 🙏

Overall, the training of DINOv3 mirror some striking aspects of brain development: late-acquired representations map onto the cortical areas with e.g. greater expansion and slower timescales, suggesting that DINOv3 spontaneously captures some of the neuro-developmental trajectory

→ Second factor: data type: Even models trained only on satellite or cellular images significantly capture brain signals — but the same model trained on standard images encodes higher all brain regions.

So what are the factors that lead DINOv3 to become brain-like?
→ 1st factor: Model size: bigger models become brain-like faster during training, reach higher brain-scores, especially in high-level brain regions.

Third, the representations of the visual cortex are typically acquired early on in the training of DINOv3.
By contrast, it requires much more training to learn representations similar to those of the prefrontal cortex.

Surprisingly, these encoding, spatial and temporal scores all emerge across training, but at different speeds.

Second, DINOv3 learns a representational hierarchy which corresponds to the spatial and temporal hierarchies in the brain.

First, we observe that, with training, DINOV3 learns representations that progressively align with those of the human brain.

To evaluate how data type, data quantity and model size each leads DINOv3 to more-or-less brain-like activation, we trained and tested several variants:

We compare the activation of DINOv3 (ai.meta.com/dinov3/), a SOTA self-supervised computer vision model trained on natural images,
to the activations of the human brain in response to the same images using both fMRI (naturalscenesdataset.org) and MEG (openneuro.org/datasets/ds0...)

Can self supervised learning help understand how the brain learns to see the world?

Our latest study, led by Josephine Raugel (FAIR, ENS), is now out:

📄 arxiv.org/pdf/2508.18226
🧵 thread below

🚨 Just over a week left to register for the #CNSP2025 Online Workshop (details in post below)! 🚨

Link to the workshop registration form: docs.google.com/forms/d/e/1F...

Our first Keynote Speaker this year will be Jean-Rémi King
@jeanremiking.bsky.social (CNRS) who leads the Brain & AI team @metaai.bsky.social. He will be giving an exciting talk on the "Emergence of Language in the Human Brain".

Mindreading with Jean Rémi-King - StarTalk Special Edition What would it take to read someone’s mind? Neil deGrasse Tyson, Chuck Nice, & Gary O’Reilly explore the science of decoding our thoughts w/ Jean-Rémi King.

Had a blast chatting with @neildegrassetyson.com on @startalkradio.bsky.social ’s podcast about how AI and neuroscience come together
startalkmedia.com/show/mindrea...

Very nice , congrats !

We are honoured to welcome @jeanremiking.bsky.social to BrainHack Donostia 2025!

Dr. King is a CNRS researcher based at the École Normale Supérieure in Paris, currently leading the Brain & AI team at Meta AI, combining expertise in cognitive neuroscience and machine learning.

Excited to be with my team at #ccn2025 this week! I’ll be presenting part of the workshop on Thursday. Come say hi!