Your Discrete Graph Diffusion Model is Secretly Memorizing

За 30 секунд

Что произошло

Researchers found that discrete graph diffusion models memorize training data when reverse trajectories are conditioned on corrupted training graphs, despite appearing to generalize.

Почему это важно

Matters for engineers building graph generation systems who need to understand whether their models truly learn generalizable patterns or exploit training data memorization.

На что обратить внимание

The memorization is conditional and hidden during normal inference, only surfacing when analyzing reverse diffusion trajectories, making it easy to miss in standard evaluation.