Reproduce “LayerDAG: A Layerwise Autoregressive Diffusion Model for Directed Acyclic Graph Generation” 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
Mufei Li, Viraj Shitole, Eli Chien, Changhai Man, Zhaodong Wang, Srinivas Sridharan, Ying Zhang, Tushar Krishna and 1 more
Researchers from Georgia Institute of Technology, Meta, and NVIDIA developed LayerDAG, a model combining layerwise autoregressive generation with diffusion models to produce large-scale, attributed Directed Acyclic Graphs (DAGs). The system generated significantly more valid DAGs on synthetic datasets and enabled ML-based surrogate models to better predict real-world system performance, improving Pearson correlation by up to 0.2 in generalization tasks.
Key insights
- A unique "layerwise tokenization" of DAGs into ordered sequences of bipartite graphs enables permutation-invariant autoregressive generation while preserving partial order.
- Integrating autoregressive layer-by-layer generation with intra-layer diffusion models allows for simultaneously capturing global directional flows and local logical dependencies.
- Intelligently varying diffusion steps per layer based on layer complexity offers an effective trade-off between the quality and computational efficiency of the generation process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.