The Forward Score Becomes the Stability Mirage
Yiwei Zhou's arXiv paper Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling tests a quiet assumption behind diffusion-model evaluation: if the learned score is accurate where the forward process usually goes, perhaps the sampler is numerically safe. The paper says that assumption is too weak.
From Score Accuracy to Sampler Stability
The paper is Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling, arXiv:2607.08757 [stat.ML]. The arXiv record lists Yiwei Zhou as author, with submission on July 9, 2026. It is categorized under Machine Learning in statistics and computer science, with numerical analysis and probability cross-listings, and the record describes the manuscript as 27 pages with 2 figures and 1 table.
The subject sounds narrow, but the governance shape is familiar. A diffusion system is trained against one distributional view, then deployed through a sampler that visits states produced by its own numerical path. The metric watches the forward marginals. The sampler lives on a reverse-time trajectory.
That puts the paper beside video cache drift records, event-video witness receipts, and visual default-prior overrides: the deployed procedure must be judged where it actually travels.
What the Paper Proves
Zhou's central distinction is between forward-marginal score error and sampler stability. The paper says score matching controls average error under forward marginals, while a discretized reverse-time sampler evaluates the learned score along the sampler's own trajectory. From that mismatch, it builds a counterexample: a smooth score field can have arbitrarily small forward-marginal L2 error, while its Euler-Maruyama discretizations converge in probability and still lose every positive moment.
The abstract and introduction make the failure sharp. The learned reverse-time process can be nonexplosive, have moments of every order, and be close to the exact reverse-time process in path-space total variation. Yet the discretized endpoints can diverge in every Wasserstein distance W_p for p >= 1. Weak convergence can coexist with exploding moment evidence.
The mechanism is not a mystical failure of diffusion. It is a blind spot in a scalar certificate. The paper places a destabilizing perturbation in a remote region with little forward-marginal mass. The continuous process almost never reaches it, so the forward-error budget and path-law comparison can look benign. A rare explicit numerical trajectory can still enter that region, and repeated updates can amplify the error faster than the entry probability decays.
The Mirage
The mirage is the belief that "accurate on the training path" means "stable under the sampling path." It is the same institutional mistake as treating a high benchmark number as a behavior receipt or treating a generated sensor reconstruction as an observation. The paper's equal-budget result says the scalar forward-marginal budget does not even reliably rank score fields by numerical stability: two fields can spend the same small budget while one sampler has grid-uniform moments and the other loses positive moments.
The fixed-architecture result matters for reviewers who would otherwise dismiss the construction as a hand-built mathematical oddity. The paper states that the obstruction persists within one fixed finite neural architecture: bounded, globally Lipschitz denoisers can have vanishing on-path score error and path-space total variation distance while Euler-Maruyama endpoints diverge in Wasserstein metrics. It also relates the construction to a fixed DiT-S/2 resource budget.
The small DiT-style experiments are evidence for the mechanism, not a production benchmark: rare numerical trajectories can show large growth, and denoiser projection can suppress that growth while ordinary trajectory errors remain small. Governance misses that case when it accepts average score accuracy, ordinary sample grids, or weak endpoint convergence alone.
What the Positive Result Means
The paper does not stop at negation. For compactly supported data, it gives a stability route through denoiser projection. If the data support is contained in a known bounded closed convex set, projecting the learned denoiser onto that set preserves pointwise accuracy, gives grid-uniform moment bounds, and yields Wasserstein convergence under mild local regularity. For image data scaled to a box, the paper connects this to clipping the predicted clean sample.
The policy lesson is precise: support assumptions are operational facts, not decorative math. A clipping or projection rule should name the support set, the denoiser parameterization, the sampler, the grid, and the failure mode it is meant to block. Otherwise "we clip the sample" becomes another folk practice with no accountable boundary.
This belongs with quantized model behavior receipts, model-collapse distance checks, and world-model coverage warnings. In all three, an aggregate success measure is weaker than a procedure that records where the model is allowed to travel and how the exceptional path is tested.
Limits and Governance
The experimental section should be read carefully. The appendix describes one fixed training seed, common evaluation seeds for paired comparisons, and Monte Carlo point estimates conditional on that run and its checkpoints. The stress protocol isolates post-entry amplification; it does not measure how often an ordinary user will naturally reach a rare region.
That limit does not weaken the warning; it keeps it honest. The paper is not saying every diffusion sampler is unstable. It is saying one common certificate is not a certificate. If an organization claims numerical stability from forward-marginal score accuracy alone, the missing evidence is now visible.
For generated media, simulation outputs, decision-support imagery, or any diffusion-backed artifact used in an institutional workflow, the receipt has to follow the sampler. It should include the reverse-time process, discretization rule, step grid, noise schedule, projection or clipping policy, support assumption, rare-trajectory stress tests, moment checks, Wasserstein checks, seed set, implementation version, and reviewer signoff.
The Receipt
A sampler-stability receipt should name the method, base model, learned-score parameterization, score loss, forward-marginal evaluation set, reverse sampler, discretization rule, grid family, step count, terminal law, support set, projection rule, clipping bounds, random seeds, weak-convergence evidence, moment diagnostics, Wasserstein diagnostics, stress construction, failure thresholds, and review path.
The Spiralist reading is simple: a forward score is not a sampler oath. The accountable object is the whole path from training metric to numerical trajectory to generated artifact.
Sources
- Yiwei Zhou, Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling, arXiv:2607.08757 [stat.ML], submitted July 9, 2026.
- arXiv experimental HTML for Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling, checked for the setup, counterexample, equal-budget result, fixed-architecture obstruction, denoiser-projection result, experiments, discussion, and protocol notes.
- arXiv API record for arXiv:2607.08757, checked for title, author, subject categories, submission date, page/figure/table comment, and version metadata.