«ASIF: Coupled Data Turns Unimodal Models to A Multimodal without Training» stood out from the crowd for the gigantic font size. Very catchy. #NeurIPS23

Exciting news! 🎉 Thanks to our generous sponsor Google DeepMind we will be offering free registration as prizes for: 🏆 Best Paper Award 🏅 Honorary Mention Make sure to submit your work here : openreview.net/group?id=NeurI… 🗓️ Deadline: Sept 20! —don’t miss out!

⚠️Unify those representations and get a free entry to NeurIPS 2024! NeurIPS Conference ⚠️ Multi-modality, model merging, representation similarity measures, and more! Paper submission deadline extended to Sept 23rd, don't miss it! 📅 Join the community 👉discord.gg/ET5NVCWwnA

#AITHYRA is born! 🎉 OeAW is founding a new institute at the intersection of #AI & #biomedicine – the AITHYRA Institute. Meet the new founding director and DeepMind Professor of Artificial Intelligence Michael Bronstein. Watch now ➡️youtu.be/pTA6DPQpxfo #boehringeringelheimstiftung

📢 Exciting news! The UniReps Blogpost Track is officially open for submissions! 🔵🔴 Submit your blog post all-year round at unireps.org/blog 👈 Want to present at the workshop poster session at NeurIPS Conference 24? Submit by **Nov 25 AoE** 📅 See you in Vancouver!🇨🇦

Forget photos — feed network parameters to your multiview reconstruction pipeline instead. Wait, what? 🤔 Not exactly that, but close: cycle consistency meets model merging! 🤝 Big shoutout to my student Donato Crisostomi for leading this exciting work—topology still shapes RepL!

Think of it like a dance-off between models 💃🤖: tune, merge, repeat, and accuracy soars up to 20%! Proud of my students Luca Zhou and Daniele Solombrino for making this happen!

Hey Demoscene can anyone make this into a 128 byte demo?

A very neat, clear, approachable and at the same time rigorous textbook, will be referring my own students to it from now on!

People from the geometry processing community probably saw this coming ;)

They say it takes two to tango👯—so how do you know if your stems are meant to dance together?🎶 Check out how, with this recent work by my amazing team and collaborators! #ICASSP2025

We found a secret recipe to eliminate task interference and boost multi-task model merging by +15%—for FREE!🧙‍♂️ In a few lines of code.🧑‍💻 💡The trick? Go low-rank AND orthogonalize. No test-time adaptation needed. All you need is a CPU and the time it takes to read this tweet✨

Buckle up for Humanity's Last Exam—I threw in a twist: cinema & electronics instead of math or code! 🍿⚡ Herr Lang🧐 would be proud of me

🎉 Excited to announce that our survey paper, "Preserving Privacy in Large Language Models: A Survey on Current Threats and Solutions," is now officially published in Transactions on Machine Learning Research ! 📚 🔗 [Read it here](openreview.net/forum?id=Ss9MT…)

*Task Singular Vectors: Reducing Task Interference in Model Merging* by Gargiulo Donato Crisostomi @ CVPR Emanuele Rodolà et al. Happy to share our work was accepted at #CVPR2025! We provide a SOTA model merging method based on the SVD of task vectors. arxiv.org/abs/2412.00081

Will present this at #CVPR2025 ✈️ See you in Nashville🇺🇸! Kudos to the team 👏 Antonio A. Gargiulo, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, Emanuele Rodolà

Just over 3 weeks left to apply for #DLG2025, the 2nd IAPR Summer School on Deep Learning on Graphs! Thanks to the support of the #IAPR we are able to provide a number of scholarships to financially support students in need. More info at sites.google.com/view/tc2-dlg-2

ARC-AGI o3 retest results are in! Takeaway: o3 (medium) is the industry leading AI reasoning system by large margin. 2X score and 1/20 cost compared to next leading chain-of-thought system as measured by ARC v1 semiprivate set scoring 57% for $1.5/task. On v2, o3 (medium)

*Mergenetic: a Simple Evolutionary Model Merging Library* by Andrea Santilli Donato Crisostomi Emanuele Rodolà Cool library to combine state-of-the-art merging techniques for LLMs with evolutionary algorithms. 🙃 arxiv.org/abs/2505.11427

Come see us at CVPR! Too bad my student Donato Crisostomi @ CVPR didn't shape the poster as a guitar🎸 -- but you'll see him tuning task vectors nonetheless! Model merging made interference-free 🤖🪡