🗒️#NeurIPS paper on tackling NP-hard problem(s) is out! How can self-training meet dynamic programming? Paper: arxiv.org/abs/2310.18672 1/2

Have you ever been left puzzled by your PPO agent collapsing out of nowhere? 📈🤯📉 We’ve all been there... We can help you with a hint: monitor your representations!💡 🚀 We show that PPO suffers from degrading representations and that this breaks its trust region 💔

Submissions to our workshop at ICML on the next "Next Generation of Sequence Modelling" are open now 🎉🎇. If you have interesting ideas on improving research related to sequence models, consider submitting them to our workshop: sites.google.com/view/ngsmworks…

How is Stable Diffusion or Dall.E able to synthesize novel images? Does this arise from the architecture? Is it the training data? 🤔 Our new paper in  #ICML2024 🚀 ICML Conference focuses on the interpolation aspect of generalization🧵 1/8

🌟 Excited to share our latest work on making diffusion language models (DLMs) faster than autoregressive (AR) models! ⚡ It’s been great to work on this with Caglar Gulcehre 😎 Lately, DLMs are gaining traction as a promising alternative to autoregressive sequence modeling 👀 1/14 🧵

Want to make the big FFN more efficient? Check out our work tomorrow! 📅 Wednesday, 11 am – 2 pm PST 📍 East Exhibit Hall A–C #2010 Stop by and explore how structured matrices can make LLM training more efficient. 🚀 Caglar Gulcehre Skander Moalla

Proud to have contributed to EvoTune — a new method that combines evolutionary search and RL fine-tuning to accelerate algorithm discovery with LLMs! 🚀 Check out claire-labo.github.io/EvoTune for more info and cool visualizations :)