Reproduce “Your Discrete Graph Diffusion Model is Secretly Memorizing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Adarsh Jamadandi, Nicolas Keriven
The authors probe training-graph memorization in a diffusion model by corrupting a graph and reversing from that state, rather than generating from an empty graph. At 256 nodes and 60% destruction, it restores 96.9% of deleted training edges, versus 7.7% for a law-only reference. The result is limited to EDGE trained on one synthetic graph.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.