α-Flow: A Unified Framework for Continuous-State Discrete Flow Matching Models
Published in arXiv preprint, 2025
Recommended citation: Chaoran Cheng, Jiahan Li, Jiajun Fan, Ge Liu. "α-Flow: A Unified Framework for Continuous-State Discrete Flow Matching Models." arXiv preprint arXiv:2504.10283, 2025. https://arxiv.org/abs/2504.10283
α-Flow unifies continuous-state discrete flow matching: existing models turn out to work on different α-representations of probabilities. Built on information geometry, α-Flow follows the canonical α-geometry of the statistical manifold, is optimal for a generalized kinetic energy, and its flow-matching loss gives a unified variational bound on the discrete negative log-likelihood; it is evaluated across several discrete generation domains.