AI · 8 min read · April 21, 2026
Three diffusion methods unified under population genetics framework
Researchers connect discrete, Gaussian, and simplicial diffusion models through Wright-Fisher theory, enabling stable cross-domain sequence generation.
Source: arxiv/cs.LG · Nuria Alina Chandra, Yucen Lily Li, Alan N. Amin, Alex Ali, Joshua Rollins, Sebastian W. Ober, Aniruddh Raghu, Andrew Gordon Wilson · open original ↗ ↗
Three separate diffusion approaches for discrete sequences share a common mathematical foundation in population genetics.
- — Discrete, Gaussian, and simplicial diffusion each model sequences differently but solve the same underlying problem.
- — Wright-Fisher population genetics model serves as unifying framework for all three methods.
- — Simplicial and Gaussian diffusion emerge as limiting cases of Wright-Fisher process at large population scales.
- — Simplicial diffusion gains numerical stability when grounded in Wright-Fisher theory instead of ad-hoc formulations.
- — Single trained model can switch between all three diffusion domains at inference time without retraining.
- — Experiments show Wright-Fisher simplicial diffusion outperforms prior simplicial methods on conditional DNA generation.
- — Multi-domain training produces models competitive with single-domain specialists across different sequence types.
- — Theory connects hyperparameters and likelihood functions across previously disconnected model families.
Frequently asked
- The Wright-Fisher model describes how allele frequencies change in a finite population over generations. It serves as a common mathematical foundation because discrete diffusion, Gaussian diffusion, and simplicial diffusion can all be derived as different parameterizations or limiting cases of the same Wright-Fisher process. This connection allows researchers to translate insights and algorithms between previously separate frameworks.