<ul><li>Researchers have introduced Masked Element-wise Learnable Diffusion (MELD) to improve the performance of Masked diffusion models (MDMs) in molecular generation.</li><li>Standard MDMs were found to severely degrade performance due to a state-clashing problem where forward diffusion causes distinct molecules to collapse into a common state.</li><li>MELD orchestrates per-element corruption trajectories using a noise scheduling network to prevent collision between distinct molecular graphs.</li><li>Experiments show that MELD significantly enhances generation quality, increasing the chemical validity of MDMs on ZINC250K benchmark from 15% to 93% and achieving state-of-the-art results in conditional generation tasks.</li></ul>

Learning Flexible Forward Trajectories for Masked Molecular Diffusion

Discover more