Marianne Arriola
@mariannearr
New work from our Cornell group with NVIDIA: fine-tuning MDMs to correct decoding errors yields up to 1.3x improvement on math+coding benchmarks!
Yair Schiff@SchiffYair · Mar 5🚨 In our paper “Learn from Your Mistakes: Self-Correcting Masked Diffusion Models” we tackle one of the most pressing issues for masked diffusion models (MDMs) → How can we enable fast parallel generation + recovery from mistakes
0:14 0 20