Efficient Adversarial Training in LLMs with Continuous Attacks Proposes a method for LLM adversarial training which does not require expensive discrete optimization steps arxiv.org/abs/2405.15589

Retraining generative models solely on their own synthetic data leads to model collapse. But what if the data was curated? With Quentin Bertrand , Joey Bose , Gauthier Gidel we show that retraining on curated data implicitly optimizes for a reward model ! 🚀 arxiv.org/pdf/2407.09499

Some exciting news coverage of iterative retraining of generative models by the NYT! nytimes.com/interactive/20… The article includes several papers including some of my co-authors from Mila - Institut québécois d'IA Quentin Bertrand Damien Ferbach Gauthier Gidel were excited to write: Broad strokes

📣

ICLR 2025 will have another blogpost track! If you have new intuitions on past work, noticed key implementation details for reproducibility, have insights into the societal implications of AI, or an interesting negative result, consider writing and submitting a blogpost.

Here is a one-minute summary of Sophie Xhonneux (Sophie Xhonneux)'s Efficient Adversarial Training in LLMs with Continuous Attacks. Come see the spotlight poster at NeurIPS Conference today: Poster Session 3 East, #4702

I'll be at #NeurIPS24 until Sunday. If you're interested in solving variational inequality problems with deep learning (e.g. min-max and projected Bellman error), come and checkout our poster on surrogate losses at the opt ml workshop. arxiv.org/abs/2411.05228

Un aperçu du poster spotlight de Damien Ferbach présenté à NeurIPS Conference, Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences:

Happy to announce "Performative Prediction on Games and Mechanism Design" was accepted at AISTATS Conference 2025, and got spotlight at HAIC(ICLR 2026 workshop) with Mehrnaz Mofakhami Fernando P. Santos Gauthier Gidel Simon Lacoste-Julien (Mila and UvA) arxiv.org/abs/2408.05146 Details below 1/9🧵

1. We often observe power laws between loss and compute: loss = a * flops ^ b + c 2. Models are rapidly becoming more efficient, i.e. use less compute to reach the same loss But: which innovations actually change the exponent in the power law (b) vs change only the constant (a)?

One of the only questions we could not solve in this paper (with Katie Everett Damien Ferbach Elliot Paquette and Courtney Paquette), should we call this figure the Hummingbird or Mario Kart Rainbow Road?

🎉Personal update: I'm thrilled to announce that I'm joining Imperial College London Imperial College London as an Assistant Professor of Computing Imperial Computing starting January 2026. My future lab and I will continue to work on building better Generative Models 🤖, the hardest