Research map / Probabilistic and scientific modeling / Variational and generative sampling
Discrete Flows: Invertible Generative Models of Discrete Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 1 autoregressive flow | Autoregressive Base · Full-rank discrete distributions | negative log-likelihood (nats): 7.7 -> 7.6; 10.7 -> 10.3; 15.9 -> 15.7 | Helped |
| Add 4 bipartite flows | Factorized Base · Full-rank discrete distributions | negative log-likelihood (nats): 8.0 -> 7.9; 11.5 -> 10.7; 16.6 -> 16.0 | Helped |
| Add 1 autoregressive flow | Autoregressive Base · Potts models | negative log-likelihood (nats): 9.27 -> 9.124; 16.66 -> 11.23; 11.64 -> 10.45; 13.58 -> 10.25 | Helped |
| Add 1 autoregressive flow | Autoregressive Base · Potts model (3 states, D=9, J=0.5) | negative log-likelihood (nats): 3.79 -> 3.79 | No clear effect |
| Add an autoregressive flow | Autoregressive Base · Addition (D=10) | negative log-likelihood (nats): 4.0 -> 0.2 | Helped |
| Add an autoregressive flow | Autoregressive Base · Addition (D=20) | negative log-likelihood (nats): 12.2 -> 4.8 | Helped |
| Increase bipartite flow depth from 1 to 4 (ablation) | Bipartite Flow · Addition (D=10) | negative log-likelihood (nats): 4.0 -> 2.58 | Helped |
| Increase bipartite flow depth from 1 to 8 (ablation) | Bipartite Flow · Addition (D=20) | negative log-likelihood (nats): 12.2 -> 5.08 | Helped |
| Use bipartite flows instead of an autoregressive LSTM | Character-level language model · Penn Tree Bank character-level test | test NLL (bpc): 1.18$^3$ -> 1.38 | Mixed |
| Use bipartite flows for parallel generation | Character-level language model · Penn Tree Bank character-level generation (288-token sequence) | generation time: 3.8 min -> 0.17 sec | Helped |
| Increase the number of bipartite flows (ablation) | text8 character-level model · text8 test | bits per character | Helped |
| Use a flexible scale transformation instead of fixing sigma=1 (ablation) | text8 character-level Bipartite Flow · text8 test | bits per character | Helped |
Closest papers by meaning
- Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows (NeurIPS 2025 · similarity 0.89)
- Flow Map Language Models: One-step Language Modeling via Continuous Denoising (NeurIPS 2026 · similarity 0.89)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.89)
- Continuous Diffusion Model for Language Modeling (NeurIPS 2025 · similarity 0.88)
- Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution (ICML 2024 · similarity 0.87)
- LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling (NeurIPS 2026 · similarity 0.87)
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.