You don't need the full corruption path. You need its snapshots.
Generalized Discrete Diffusion from Snapshots (arXiv:2603.21342): discrete diffusion models — which corrupt and reconstruct data in finite state spaces (text, molecular structures, categorical data) — traditionally require tracking the full forward corruption process. GDDS replaces the full path with snapshots: latent states sampled at discrete times along the corruption trajectory. The simplified ELBO derived from snapshots rather than paths enables efficient training with flexible corruption dynamics.
Reader note (2026-11-17): Checked against the full text of arXiv:2603.21342 (Zekri, Uscidda, Boullé and Korba; v1), which I had earlier held as a weak read. The framework, the uniformization of the forward process and the snapshot ELBO match. Three points differ. First, the essay says the shortcut is not an approximation and that the path is redundant and the snapshots sufficient. The paper's own Proposition 4.4 decomposes the gap between the snapshot and the path-wise objectives into an intrinsic information path gap, which it states is non-negative (the path carries information the snapshot does not), plus a calibration term; it reports that empirically the net gap is often negative because the snapshot objective improves calibration enough, and its snapshot objective is an evidence lower bound obtained with Jensen's inequality. So the paper presents a trade in which the path loses information and calibration gains, not redundancy. Second, the performance claim: the essay says competitive with autoregressive models on large-vocabulary tasks; the abstract says the experiments suggest the framework outperforms existing discrete diffusion methods and beats autoregressive models for the first time at this scale, and the text reports it outperforming the autoregressive baseline on Text8 and being able to outperform the matched baseline on OpenWebText. Third, the scale: those comparisons are at about 96 million non-embedding parameters for the diffusion models against about 89 million for the autoregressive baseline, under matched compute, and the essay says surpasses where the abstract says suggest.
The flexibility matters. Standard discrete diffusion constrains the corruption to specific noise schedules — typically uniform or absorbing. GDDS uses uniformization to separate the corruption rate from the corruption dynamics, allowing arbitrary corruption processes across large state spaces. The forward process can be any continuous-time Markov chain. The reverse process only needs to match the marginal distributions at snapshot times, not reconstruct the full path between them.
The result: surpasses existing discrete diffusion methods in training efficiency and generation quality, competitive with autoregressive models on large-vocabulary tasks.
The structural principle: the full trajectory contains information the model doesn't need. Between any two snapshots, the exact corruption path is a nuisance variable — averaging over it gives the same generative quality with less computational cost. The shortcut is not an approximation. It's a recognition that the information content of the process lives at discrete checkpoints, not in the continuous path between them. The path is redundant; the snapshots are sufficient.