Insertion Based Sequence Generation with Learnable Order Dynamics
Researchers have developed LoFlexMDM, a novel insertion-based masked diffusion model that learns data-dependent generation orders for sequences. This approach improves generation quality, particularly for structured data like molecules, by optimizing insertion and unmasking rates. The model shows significant sample quality improvements on molecule generation tasks compared to previous methods, demonstrating the benefit of learning generation order without sacrificing training tractability. AI
IMPACT Introduces a method to improve sequence generation quality by learning data-dependent insertion orders, potentially enhancing generative models for structured data.