EuroBERT is going to Conference on Language Modeling 2025! Can’t wait to be in Montreal with Hippolyte Gisserot-Boukhlef and Duarte Alves to see all the great research everyone’s bringing!

"Reasoning models improve performance" - even when reasoning traces are distilled through SFT to smaller student models as Qwen showed. This paper attempts to quantify on what tasks, model sizes, and to which extent this holds true w.r.t. instruct tuning. 🧪

🚨 We release ModernVBert - a small bidirectional ModernBert encoder trained to process image inputs alongside text. When finetuned for document retrieval, it matches the original ColPali performance on ViDoRe with almost 10x fewer parameters (250M). (1/N)

I'm heading soon to Montreal for Conference on Language Modeling ! Our lab is presenting the following 5 papers: 🧵

5) EuroBERT: Scaling Multilingual Encoders for European Languages w/ Nicolas Boizard Hippolyte Gisserot-Boukhlef Duarte Alves Ayoub Hammal Caio Fanny Jourdan @COLM2025 Manuel Faysse Maxime Peyrard Patrick Fernandes Ricardo Rei Pierre Colombo tomaarsen - Poster session 5, Thu Oct 9, 11:00 AM – 1:00 PM

Come along on Thursday 9th if you want to find out more about EuroBERT or simply chat with the project members 🙌

I am at #COLM2025 this week. 👋 Feel free to contact me if you would like to discuss interpretability and evaluation of LLMs. And don't miss the Poster session 5 (Thu Oct 9, 11AM – 1PM) where EuroBERT ( lnkd.in/g_RXSjQX ) will be presented by Hippolyte Gisserot-Boukhlef & Nicolas Boizard !

COLM 2025 was once of the best conferences I can recall ever attending. I heard this also from many in coffee chats and dinners. Thank you to all that put it together. I look forward to year three.

🚀 It’s official: TowerVision is live! Fully open-source multilingual Vision-Language Model 🔗 Paper: arxiv.org/abs/2510.21849 🌐 Project page: guilhermeviveiros.github.io/TowerVision.io/ 🤗 Models: huggingface.co/collections/ut… #OpenSource #Multilingual #VLM

What a great work by Carlos Miguel Patiño and the Hugging Face teams! So happy to see that ULD Loss inspired the idea of their GOLD method and that they solved the early sequence and vocabulary alignment issues we faced. This recognition is the greatest I could have hoped for when I first

ViDoRe V3: the last effort of my PhD arc towards pushing for better document retrieval. We focused a lot on architectures (ColPali, ColQwen, ModernVBert) and evaluations always took the back seat - costly, less fun, hard to get right the first time... Here we flip this around ⤵️