Source
Sahoo, Deschenaux, Gokaslan, Wang, Chiu and Kuleshov (2025), LIT-479 §3 — read as NOTE-228.
What it explains
| practice | what it says to do | what this says it is |
|---|---|---|
| SOTA-289 | for few-step generation, use uniform-state diffusion with consistency distillation | not a lucky transplant but a licensed one: the discrete process has a continuous preimage, so the continuous literature’s methods have somewhere to land |
The account
Run a Gaussian diffusion on R^K and map each latent to the one-hot vector of
its largest coordinate. Two things then have to be shown, and both are.
The marginals match. The argmax of a Gaussian latent has a closed-form
probability mass function, and it is exactly the marginal of a uniform-state
discrete diffusion — with the noise parameter reparameterized by a
diffusion transformation operator built from the standard normal CDF. The
two processes are different Markov chains; what the map relates is their
marginal distributions.
The dynamics match. The discretized marginal evolves under the linear ODE that characterizes a discrete diffusion process. Satisfying that ODE is necessary and sufficient, so the discretized Gaussian process is a uniform-state discrete diffusion rather than something that resembles one.
Both halves are derived in LIT-479 §3, which builds the marginal transformation from the standard normal CDF, derives the transition ODE, and verifies one of the two numerically. It is the paper that then builds Duo on the result.
The consequence is a licence. Techniques developed for Gaussian diffusion —
noise schedules, variance reduction, consistency distillation — are not being
applied to discrete diffusion by analogy; they are being applied to a process
that has a Gaussian representation. The paper cashes this twice, with a
curriculum that anneals the argmax through a tempered softmax and with a
discrete adaptation of consistency distillation.
And the asymmetry is the point. No comparable map exists for absorbing-state (masked) diffusion. The source’s conclusion says so directly, and it is the clearest statement of why the two discrete families should be expected to diverge in what can be borrowed for them, whatever their relative perplexities today.
Why Active
Because it is a derivation with both halves shown, and one of them checked
numerically as well. The status is about the account, not about the method’s
standing: SOTA-289 is Proposed and unreplicated, and an
explanation being correct is not evidence that a practice is worth adopting —
the split ADR-031 exists to keep visible.
What this does not say
It does not say uniform-state diffusion is better. Masked diffusion is ahead on likelihood in the source’s own tables, on 6 of 7 zero-shot datasets and on both training corpora. The duality is a claim about what can be borrowed, not about what currently wins.
It does not carry the empirical result it was used to obtain. The source’s ablation splits its three-point perplexity improvement over the previous uniform-state model roughly evenly between a Rao-Blackwellized ELBO — which has nothing to do with this account — and the duality-derived curriculum. The theorem earns about 1.3 of those 3 points.
And the curriculum it licenses is not itself a bound. Feeding the denoising model a continuous latent while the objective is defined for a discrete process makes the training loss an invalid NELBO except in the limit. The source says this and evaluates with the proper discrete bound instead. So the theorem justifies the shape of the training procedure, not its status as a likelihood bound.
Two demonstrations is not the general claim. That Gaussian technique transfers is shown for a curriculum and for consistency distillation. Whether the bridge carries the rest of that literature is the open question, and the value of the account depends on the answer.