Heavy-Tailed Flow Matching via Random Clocks
The paper replaces Gaussian flow-matching sources with random-clock Gaussian mixtures to better fit heavy-tailed data.
HTFM treats heavy-tailed sources as Gaussian distributions conditioned on random clock paths, then marginalizes over the clock. The authors use truncated logsignature features so the vector field can respond to the realized clock path with little added cost. In tests on imbalanced alpha-stable mixtures, CIFAR10-LT, and HRRR weather fields, they report better mode coverage, sample quality, and tail-statistic recovery than Gaussian flow matching and other heavy-tailed baselines. The same setup also lets users tune generated tail heaviness by changing the clock law or tail parameter. ArXiv · AI/CL/LG's note
HTFM treats heavy-tailed sources as Gaussian distributions conditioned on random clock paths, then marginalizes over the clock. The authors use truncated logsignature features so the vector field can respond to the realized clock path with little added cost. In tests on imbalanced alpha-stable mixtures, CIFAR10-LT, and HRRR weather fields, they report better mode coverage, sample quality, and tail-statistic recovery than Gaussian flow matching and other heavy-tailed baselines. The same setup also lets users tune generated tail heaviness by changing the clock law or tail parameter. ArXiv · AI/CL/LG's note
score 4