AI Scientist - Safety & Security @MistralAI | PhD from @MetaAI and @Polytechnique I confront equations and inequalities💡 my tweets reflect my own views only.
Paris' suburb
Joined December 2010
- Tweets3.4K
- Following2.3K
- Followers773
- Likes3.1K
Wassim (Wes) Bouaziz retweeted
Chinchilla writes loss as a sum of two independent terms: model size and data never interact. Kaplan's original law coupled them.
But interaction has an impact, dropping it bends predictions exactly where you extrapolate. Skaling, our new law, puts it back with one exponent. 1/8
Wassim (Wes) Bouaziz retweeted
🧵How to combine code correctness and efficiency objectives in online RL?
-> we break the correctness-efficiency Pareto frontier!
+125% relative improvement in CWM32B compared to standard RLVR (13.7 ->30.9 pass@1@top30%)
+150% with Qwen32B
Paper: arxiv.org/abs/2607.25970
Wassim (Wes) Bouaziz retweeted
In the last project, we study transformer training in a controllable and interpretable setting, providing a visual sandbox to monitor tokens during gradient descent. Thanks to @_Vassim and Vivien Cabannes for the really nice collab.
📜openreview.net/pdf?id=eJTS0b…
Wassim (Wes) Bouaziz retweeted
🧵 For 2 RL checkpoints trained differently, you can just weight extrapolate them and it works!
Bonus: these extrapolated checkpoints are complementary policies
-> Get exploration and diversity for free
-> Better inference scaling when ensembling
Paper: arxiv.org/abs/2605.28751
arxiv.org/abs/2605.28751
Now many studies try to do extrapolation through model merging. arxiv.org/abs/2605.26484
Excited to see the recent discussions around trait transfer in LLMs! 🦉
It validates the idea of Indirect Data Poisoning we introduced in our Winter Soldier paper (arxiv.org/abs/2506.14913) which predates the Subliminal Learning work.
It's important to connect these lines of work.
I'm particularly proud of the Winter Soldier project, being my last PhD project, and it's great to see the broader community engaging with these ideas 🤗
Data are not just passive inputs, and training on them could lead to similar risks as running an untrusted program 🛡️💻
We have tons of terminal apps but all rely on the same outdated multiplexer...
Use a terminal that supports tmux's control mode (`tmux -CC`) and you'll never have to learn a single tmux "short"cut again
Wassim (Wes) Bouaziz retweeted
✨Vision Transformer finetuning benefits from non-smooth components
🔍Our new paper shows that high-plasticity transformer modules adapt better during finetuning.
🍕 Whether you prefer theory or experiments, we hope you'll find something you like in this work.
Details below 🧵
Wassim (Wes) Bouaziz retweeted
My first PhD paper is out! 🎓
"What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?"
tl:dr: JEPA-WMs for robotics: learn dynamics on top of visual encoders, optimize actions towards goal 👇
w/ @JimmyTYYang1, Jean Ponce, @AdrienBardes, @ylecun
Big milestone 🎓✨
I’ve successfully defended my PhD thesis at @Polytechnique in collaboration with @AIatMeta !
"Towards Secure and Trustworthy Machine Learning: From Data Poisoning to Ownership Verification"
Grateful to my advisors, jury, and everyone who supported me 🙏
Open review leak has actual implications... It's safe to assume anonymity is compromised this year for @iclr_conf
Wassim (Wes) Bouaziz retweeted
I am happy to share the work of our team. The outcome of a collaborative effort, by a joyful group of skilled and determined scientists and engineers! Congrats to the team on this amazing milestone!
Introducing DINOv3: a state-of-the-art computer vision model trained with self-supervised learning (SSL) that produces powerful, high-resolution image features. For the first time, a single frozen vision backbone outperforms specialized solutions on multiple long-standing dense prediction tasks.
Learn more about DINOv3 here: ai.meta.com/blog/dinov3-self…
Wassim (Wes) Bouaziz retweeted
Say hello to DINOv3 🦖🦖🦖
A major release that raises the bar of self-supervised vision foundation models.
With stunning high-resolution dense features, it’s a game-changer for vision tasks!
We scaled model size and training data, but here's what makes it special 👇
Peer review ML conferences:
> Only "top" X% get in
> Jobs, grants, bonuses hinge on it
> No penalty for bad-faith reviews
> No cost for flooding submissions
Who could have seen this going wrong? 🤭
Wassim (Wes) Bouaziz retweeted
🚀New paper alert! 🚀
In our work @AIatMeta we dive into the struggles of mixing languages in largely multilingual Transformer encoders and use the analysis as a tool to better design multilingual models to obtain optimal performance.
📄: arxiv.org/abs/2508.02256
🧵(1/n)
Wassim (Wes) Bouaziz retweeted
If you are attending ACL2025 join our oral presentation! Happening at 15:00 in room 1.86 🙂
Very happy that our paper MEXMA has been featured on the bundle release by @AIatMeta 🎉
paper: arxiv.org/abs/2409.12737
code: github.com/facebookresearch/…
Model: huggingface.co/facebook/MEXM…
Checkout the full bundle here: ai.meta.com/blog/fair-news-s…
Wassim (Wes) Bouaziz retweeted
🚀 We are happy to organize the BERT²S workshop @NeurIPSConf 2025 on Recent Advances in Time Series Foundation Models.
🌐 berts-workshop.github.io
📜Submit by August 22
🎓Speakers and panelists: @ChenghaoLiu15 Mingsheng Long @zoe_piran @danielle_maddix @atalwalkar @qingsongedu
