Research map / Probabilistic and scientific modeling / Variational and generative sampling
Categorical Normalizing Flows via Continuous Transformations
Phillip Lippe, Efstratios Gavves
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 |
|---|---|---|---|
| add largest-sub-graph postprocessing to generated molecule graphs (ablation) | GraphCNF · Zinc250k | validity: 83.41% -> 96.35% | Improved |
| use parallel GraphCNF generation instead of GraphAF's autoregressive generation | molecule generation · Zinc250k | validity: 68% -> 83.41% | Improved |
| use GraphCNF instead of GraphNVP | molecule generation · Zinc250k | validity: 42.60% -> 83.41% | Improved |
| use GraphCNF instead of the best autoregressive node ordering | graph coloring · 10≤|V|≤20 | validity: 93.41% -> 94.56% | Improved |
| use GraphCNF instead of the best autoregressive node ordering | graph coloring · 25≤|V|≤50 | validity: 71.32% -> 66.80% | Hurt |
| replace logistic mixture coupling layers with affine coupling layers (ablation) | GraphCNF · graph coloring | validity: 94.56% -> 93.90%; 66.80% -> 65.78% | Hurt |
| replace the mixture encoding with category-conditioned linear flows (ablation) | CNF · Set shuffling and set summation | bits per categorical variable: 2.78 -> 2.78; 2.24 -> 2.25 | No clear effect |
| replace the mixture encoding with variational encoding across categorical variables (ablation) | CNF · Set shuffling and set summation | bits per categorical variable: 2.78 -> 2.79; 2.24 -> 2.25 | No clear effect |
| use factorized decoding rather than a joint decoder | categorical normalizing flows · text8 and Wikitext103 | language-modeling performance | Improved |
| increase the decoder-loss weight for Latent NF | Latent NF · Wikitext103 | bits per dimension: 10-20 | Hurt |
| use a larger latent dimensionality (ablation) | CNF encoding · categorical datasets | performance | No clear effect |
Closest papers by meaning
- Variational Flow Matching for Graph Generation (NeurIPS 2024 · similarity 0.87)
- Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation (NeurIPS 2026 · similarity 0.86)
- Causally Consistent Normalizing Flow (AAAI 2025 · similarity 0.85)
- Latent Bayesian Optimization via Autoregressive Normalizing Flows (ICLR 2025 · similarity 0.85)
- Smooth Interpolation for Improved Discrete Graph Generative Models (ICML 2025 · similarity 0.85)
- DeFoG: Discrete Flow Matching for Graph Generation (ICML 2025 · similarity 0.85)
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.
