Pablito

💫 Introducing NeuralSet: a simple, fast, scalable Python package for Neuro-AI 📦 pip install neuralset 📄 kingjr.github.io/files/neuralse… 🔍 facebookresearch.github.io/neuroai/neural… Supports 🧠 fMRI, EEG, MEG, ECoG, spike… preprocessing 💬 text 🔊 audio ▶️ video 🏞️ image… embeddings 🧵 Details👇

7/ For more details, refer to the paper. arxiv.org/pdf/2508.03211

6/ In conclusion, we find that linear probes are sensitive to linear distance and syntactic depth. This suggests that a linear readout might not be enough to capture complex syntactic patterns from the LLM's activation spaces.

Happy to share our new work accepted at #COLM2025 Probing syntax in LLMs: successes & remaining challenges 🔗arxiv.org/pdf/2508.03211 Joint work with Emmanuel Chemla, @JeanRemiKing & @lakretz Check out the poster Wednesday afternoon (Poster #40) Follow the 🧵for the details!

⚡️Check out our workshop tomorrow at @lpiparis_, with great speakers (@gabrielpeyre, @stephanedascoli, @TonyBonnaire + many more) covering Theory and Applications of Generative AI as well as Connexions with neuroscience 🧠 And there's food 🍰 ➡️ genai-conference-website.vercel.app

@rocioHF15 Hola, estoy interesado

Two new studies from our team we're particularly happy about Study 1: Brain-to-Text Decoding: ai.meta.com/research/publi… Study 2: From Thought to Action: ai.meta.com/research/publi… Blog: ai.meta.com/blog/brain-ai-…

Learn more about the Polar Probe in our full paper 📚 🔍 Link: arxiv.org/abs/2412.05571

We created a controlled dataset with 3 levels of nesting and analyzed Polar Probe embeddings in PCA space, revealing: 👉 Words have fixed positions by syntactic role, regardless of meaning or number of nestings. 👉 Nestings form hierarchical branches from the Main Phrase.

🎉 Excited to share our new paper: "A polar coordinate system represents syntax in large language models" accepted at #NeurIPS2024. Joint work with @stephanedascoli Emmanuel Chemla @lakretz and @JeanRemiKing. Check out the thread below for all the details! 🧵

@gregdeletang Oh right! I was thinking more like inputting a single vectorial representation for the whole number indeed.

@itspablit0 I'd say probably easy if a and b are given as raw inputs, but as a sequence of digits, a lot harder and I don't see how. But yes then you apply the same layer again and again a la Universal Transformer and you get it :)

Anyone knows a neural model that can sum and/or multiply numbers with infinite digits? (say up to 1B digits, it's probably enough)