*Localizing Task Information for Improved Model Merging and Compression* by Dimitriadis Nikos @ ICLR Ke Wang Guillermo Ortiz-Jiménez François Fleuret Pascal Frossard Binary masks allow to approximately retrieve a task vector from a multi-task model for better performance. arxiv.org/abs/2405.07813

Struggling to integrate structural constraints in graph generative models? 🧩 Our new paper "Generative Modelling of Structurally Constrained Graphs" presents ConStruct to your rescue! 🚀 📜: arxiv.org/abs/2406.17341 🧵1/9

Good morning ICML🇦🇹 Presenting today "Localizing Task Information for Improved Model Merging and Compression" with Ke Wang Guillermo Ortiz-Jiménez François Fleuret Pascal Frossard. Happy to see you at poster #2002 from 11:30 to 13:00 if you are interested in model merging & multi-task learning!

Super cool poster «Localizing Task Information for Improved Model Merging and Compression» by Ke Wang Dimitriadis Nikos @ ICLR François Fleuret x.com/nikdimitriadis…

It was fun to present our work at GRaM Workshop at ICML 2024 yesterday :) Thanks to everyone who stopped by to discuss, and thanks to the organizers for making such an inspiring workshop happen!

Are you interested in graph generation, from molecular discovery 🧪 to social networks 🌐? You’ll love DeFoG 🌬️😶‍🌫️, our new framework that delivers state-of-the-art performance in diverse graph generation tasks with unmatched efficiency! 🤩 📄: arxiv.org/abs/2410.04263 🧵1/9

LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging Scales parameter updates linearly based on their layer depth. Shallow layers keep to pre-trained values to preserve general features. Deeper layers retain task-specific representations Links below

Excited to share this work with Ke Wang, Alessandro Favero, Guillermo Ortiz-Jiménez, François Fleuret,Pascal Frossard! Learn more in our paper and check out the code here: 📜: arxiv.org/abs/2410.17146 💻: github.com/wang-kee/LiNeS 🧵 11/11

Amazed by all these model merging methods? Check out LiNeS, a simple post-training technique that reduces task interference and enables merging at scale by building on the idea of hierarchical representations in deep networks!

Can we edit models after fine-tuning to mitigate catastrophic forgetting? Check out our latest work - LiNeS 📈for this and more application scenarios in model merging: improving multi-task model merging, OOD generalization, model soups and even merging reward policies via RLHF!

TL;DR: Task arithmetic but downscaling changes of the early layers so that subsequent layers are not fed inputs too different from what they used to. The idea is that later layers are more symbolic / semantic so it hurts less to sum them as-is. Works well!

Surprisingly, if you downscale the mixing strength of early layers you massively reduce interference in model merging 🔥 LiNeS works out of the box and gives huge boosts to model merging 💪 Another great work led by Ke Wang and Dimitriadis Nikos @ ICLR

Mitigating racial bias from LLMs is a lot easier than removing it from humans! Can’t believe this happened at the best AI conference NeurIPS Conference We have ethical reviews for authors, but missed it for invited speakers? 😡

Variational Flow Matching goes Riemannian! 🔮 In this preliminary work, we derive a variational objective for probability flows 🌀 on manifolds with closed-form geodesics. My dream team: Floor Eijkelboom Alison Erik Bekkers 💥 📜 arxiv.org/abs/2502.12981 🧵1/5

Excited to be in Singapore next week for #ICLR2025! 🇸🇬 DM if you want to chat about geometric deep learning and/or generative models on manifolds, or just to enjoy some nice specialty coffee ☕️ I don't look like Barbie but my poster does 🪩😎

Heading to Singapore for ICLR! 🇸🇬 Excited to connect and chat about anything LLMs, model merging and editing, multi-task learning, feel free to DM if you’re around! We’ll be presenting our paper LiNeS: lines-merging.github.io this Saturday, feel free to drop by!

Happy to share that DeFoG: Discrete Flow Matching for Graph Generation will be showcased as a Spotlight Poster at #ICML2025 ! -> Explore the paper: arxiv.org/abs/2410.04263 -> Open-source code: github.com/manuelmlmadeir… Looking forward to your feedback on our repository!

Happy to release the code for DeFoG (github.com/manuelmlmadeir…), accepted as an ICML 2025 Spotlight Poster, with Manuel Madeira! → Efficient training & sampling → Wide dataset support & extensible dataloader → Hydra-powered CLI & WandB → Docker/Conda setup