Research map / Probabilistic and scientific modeling / Variational and generative sampling
A Time-Reparameterized Cumulative Intensity Extrapolation Sampler for Discrete Flow Matching
Feiyang Fu, Hehe Fan
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 schedule-based time reparameterization (ablation) | Euler tau-leaping (uniform DFM) · Text generation, uniform DFM, NFE=8 | GPT-2 perplexity: 439.94 -> 376.06 | Improved |
| add cumulative intensity extrapolation (ablation) | Euler tau-leaping (uniform DFM) · Text generation, uniform DFM, NFE=8 | GPT-2 perplexity: 439.94 -> 275.87 | Improved |
| combine TR and CIE (ablation) | Euler tau-leaping (uniform DFM) · Text generation, uniform DFM, NFE=8 | GPT-2 perplexity: 275.87 -> 209.84 | Improved |
| use nonuniform physical-time grid without reparameterized extrapolation (ablation) | Euler tau-leaping (uniform DFM) · Text generation, uniform DFM, NFE=8 | GPT-2 perplexity: 439.94 -> 401.32 | Improved |
| replace Euler with TR-CIE | Masked DFM · Text generation, masked DFM, NFE=8 | GPT-2 perplexity: 495.25 -> 252.15 | Improved |
| replace Euler with TR-CIE | Masked DFM · Text generation, masked DFM, NFE=256 | GPT-2 perplexity: 151.16 -> 135.92 | Improved |
| replace Euler with TR-CIE | FUDOKI · GenEval, NFE=8 | accuracy (%): 60.67 -> 64.27 | Improved |
| TR-CIE versus theta-RK2 at lowest budget | TR-CIE (FUDOKI) · GenEval, NFE=4 | accuracy (%): 54.32 -> 53.05 | Hurt |
| replace Euler with TR-CIE | FUDOKI · CLIP score, NFE=4 | CLIP score: 0.311 -> 0.321 | Improved |
| increase lower clamp threshold epsilon_0 (ablation) | TR-CIE · Text generation, uniform DFM | GPT-2 perplexity | Hurt |
| replace Euler with TR-CIE | DFM countdown · Countdown task | error rate | Improved |
| replace Euler with TR-CIE (unigram entropy) | Masked DFM · Text generation, masked DFM | unigram entropy: 8.03 -> 7.92 | No clear effect |
Closest papers by meaning
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.90)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.89)
- Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.88)
- Flow matching achieves almost minimax optimal convergence (ICLR 2025 · similarity 0.88)
- Drift Flow Matching (NeurIPS 2026 · similarity 0.88)
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (ICLR 2025 · 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.
