AI TrackerFrontier desk · AEST
Back to news
normalaitech

Your Discrete Graph Diffusion Model is Secretly Memorizing

arXiv cs.AI · 10 Oct 2026, 02:04 AEST

Discrete graph diffusion models (GDMs) have been found to exhibit memorization, challenging previous assumptions about their performance. This conditional memorization occurs when the model is conditioned on a corrupted training graph, leading to a significant generalization gap.

Key facts

Read aloud

Discrete graph diffusion models GDMs have been found to exhibit memorization, challenging previous assumptions about their performance. This conditional memorization occurs when the model is conditioned on a corrupted training graph, leading to a significant generalization gap. Discrete graph diffusion models GDMs corrupt the graph structure by adding noise independently to nodes and edges under a Markovian noise model. In this work, we challenge this view and show for the first time that graph diffusion models indeed exhibit memorization. We find a strong train-data bias: the model recovers a corrupted training graph almost perfectly compared to a held-out test graph. The onset of memorization always precedes the emergence of community structure in the generated graphs.

Sources in this cluster