Research map / Graph learning / Graph self-supervised learning
Bures-Wasserstein Flow Matching for Graph Generation
Keyue Jiang, Jiahao Cui, Xiaowen Dong, Laura Toni
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace linear interpolation with Bures-Wasserstein interpolation in flow matching | BWFlow · Planar graphs | V.U.N. and A.Ratio: V.U.N. 77.5 -> 84.8, A.Ratio 3.5 -> 2.4 vs DeFoG | Improved |
| use BW interpolation on tree graphs | BWFlow · Tree graphs | V.U.N.: 81.5 vs 84.5 (Cometh) | No clear effect |
| use BW interpolation on tree graphs | BWFlow · Tree graphs | A.Ratio: 1.9 -> 1.3 vs DeFoG | Improved |
| use BW interpolation, with 30 instead of 1k sampling steps | BWFlow · Planar (30 sampling steps) | V.U.N.: 72.0 -> 77.0 vs DeFoG-2; 24.5 for DeFoG-1 | Improved |
| use BW interpolation, with 30 sampling steps | BWFlow · SBM (30 sampling steps) | V.U.N.: 47.5 -> 52.0 vs DeFoG-2 | Improved |
| add path manipulation (time distortion, target guidance) to linear-interpolation flow (ablation) | DeFoG · Planar (30 sampling steps) | V.U.N.: 24.5 -> 72.0 | Improved |
| use BW interpolation instead of linear, harmonic or geometric interpolation (ablation) | BWFlow · Plain graphs | V.U.N. | Improved |
| use BW interpolation instead of linear interpolation (ablation) | BWFlow · Planar training convergence | convergence speed | Improved |
| use harmonic or geometric interpolation (ablation) | BWFlow · SBM | path validity | Hurt |
| use BWFlow for 3D molecule generation without bond types | BWFlow · QM9 with h | Mol.Stab.: 93.98 -> 97.84 vs MiDi | Improved |
| use BWFlow for 3D molecule generation without bond types | BWFlow · GEOM-DRUGS with h | Connected (%): 79.15 -> 73.53 vs MiDi | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DeFoG — DeFoG: Discrete Flow Matching for Graph Generation (ICML 2025) · better on 4 of 6 shared results
- DiGress — DiGress: Discrete Denoising diffusion for graph generation (ICLR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DeFoG: Discrete Flow Matching for Graph Generation (ICML 2025 · similarity 0.91)
- Smooth Interpolation for Improved Discrete Graph Generative Models (ICML 2025 · similarity 0.90)
- Variational Flow Matching for Graph Generation (NeurIPS 2024 · similarity 0.90)
- Graph Generation with Diffusion Mixture (ICML 2024 · similarity 0.89)
- Flowette: Flow Matching with Graphette Priors for Graph Generation (NeurIPS 2026 · similarity 0.89)
- Discrete-state Continuous-time Diffusion for Graph Generation (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
