Research map / Probabilistic and scientific modeling / Variational and generative sampling
Tail Annealing for Heavy-Tailed Flow Matching
Jean Pachebat
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 |
|---|---|---|---|
| apply soft-log transform before flow matching (Log-FM vs TTFfix) | flow matching · Gumbel+Gaussian, tail index 1.5, d=20 | median W1 (Pareto margins): 0.963 -> 0.451 | Improved |
| apply soft-log instead of arcsinh transform | Log-FM · Gumbel+Gaussian, tail index 1.5, d=20 | median W1 (Pareto margins): 0.550 -> 0.451 | Improved |
| use Log-FM instead of gTAF at high dimension | gTAF · Gumbel+Gaussian pooled, d=100 | median W1 (Pareto margins): 0.704 -> 0.233 | Improved |
| use Log-FM for Gaussian margins vs TTF | TTF · Gumbel+Gaussian pooled, d=50 | median W1 (Gaussian margins): 0.030 -> 0.080 | Hurt |
| add Hill-gated per-coordinate transform (leave light-tailed margins untouched) (ablation) | Log-FM uniform · Gumbel d=20 tau=0.5, tail index 1.5 | median W1 (Gaussian margins): 0.074 -> 0.043 | Improved |
| add Hill-gated per-coordinate transform (ablation) | Log-FM uniform · Gumbel d=20 tau=0.5, tail index 1.5 | median W1 (Pareto margins): 0.449 -> 0.468 | No clear effect |
| add Hill gating at tail index 2.5 (ablation) | Log-FM uniform · Gumbel d=20 tau=0.5, tail index 2.5 | median W1 (Pareto margins): 0.068 -> 0.084 | Hurt |
| add sampling clamp c=5 (ablation) | Log-FM · Gumbel d=20 tau=0.5, tail index 1.5 | relative W1 Pareto change (%): -2.0% | No clear effect |
| increase Euler solver steps from 10 to 500 (ablation) | Log-FM · Gumbel d=20 tau=0.5, tail index 1.5 | W1 (Pareto margins): 0.521 -> 0.588 | No clear effect |
| use Log-FM instead of TTFfix on Student-t data | TTFfix · Student-t benchmark, d=10, nu=1.5 | W1: 1.78 -> 0.63 | Improved |
| use Log-FM on lighter-tailed Student-t (nu=5) | TTFfix · Student-t benchmark, d=10, nu=5 | W1: 0.12 -> 0.14 | Hurt |
| use Log-FM instead of TTFfix on real finance data | TTFfix · Fama-French 5 | W1: 0.449 -> 0.133 | Improved |
Closest papers by meaning
- Heavy-Tailed Flow Matching via Random Clocks (NeurIPS 2026 · similarity 0.89)
- Heavy-Tailed Diffusion Models (ICLR 2025 · similarity 0.88)
- Flexible Tails for Normalizing Flows (ICML 2025 · similarity 0.87)
- Heavy-Tailed Diffusion with Denoising Levy Probabilistic Models (ICLR 2025 · similarity 0.85)
- Slimming the Fat-Tail: Morphing-Flow for Adaptive Time Series Modeling (ICML 2025 · similarity 0.84)
- SSE-SAM: Balancing Head and Tail Classes Gradually Through Stage-Wise SAM (AAAI 2025 · similarity 0.84)
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.
