The Pulse
Apple Paper Finds Confidence Rankings Can Distort Diffusion Outputs
Apple’s paper on discrete diffusion finds that confidence-ranked parallel token writes can produce the wrong output distribution when tokens depend on one another. In a synthetic task, the authors measured a 29-fold gap over the sampling-no
AI.info Team · Written with AI ·
A paper submitted to arXiv on September 17 argues that a common shortcut in discrete diffusion—writing several tokens at once according to their confidence scores—can distort the distribution of generated sequences. The study, “Limits of Confidence in Diffusion,” tests the problem on a synthetic task, where its authors measure a 29-fold gap over the sampling-noise floor even when per-sample metrics reach 1.0. The result points to a mismatch between judging tokens one by one and preserving the relationships among them.
Confidence scores see tokens, not their dependencies
Discrete diffusion generates a sequence by filling in multiple token positions over successive steps. In confidence-ranked methods, a sampler uses per-position probability distributions both to choose what value to place at each position and to decide which positions to fill first. That approach can be efficient, but it treats the selected positions as independent draws.
“We show that a step matches the training distribution only when the positions it writes are conditionally independent given the tokens already fixed.”
Russ Webb, Amitis Shidani, Alice Bizeul and Dan Busbridge, authors of “Limits of Confidence in Diffusion”
The authors’ point is that individually plausible tokens can form a combination that the training data would never contain. Per-position predictions alone cannot reveal whether a group of tokens is dependent: two joint distributions can share the same individual token probabilities while allowing different combinations. A confidence score for each token therefore cannot certify that writing several of them together will preserve the original distribution.
ScanAndAdd exposes the ordering problem
To make the dependencies measurable, the researchers built ScanAndAdd, a task with a known joint distribution. Each example contains nine values and a sequence of commands, including two additions; a simulated head moves across the values, and the commands determine the final answer. Because the answer depends on both the values and command order, several fields are linked even when their individual token predictions look sound.
In the task, confidence ranking writes the answer digits before the value positions. That order matters: once an answer is fixed, it becomes dependent on values and commands that have not yet been written. The paper reports that every group of two or more undetermined positions written together by a confidence ranking is dependent in this setup, so independent sampling can produce combinations outside the task’s valid distribution.
Perfect-looking samples still miss the distribution
The paper’s headline result separates sample-level scores from distribution-level fidelity. On ScanAndAdd, per-sample metrics reach 1.0, yet the generated distribution’s total-variation distance is 29 times the sampling-noise floor. In other words, individual outputs can satisfy the measured checks while the collection of outputs still differs from the training distribution.
The comparison also shows that write order matters. The authors report that a hand-specified order which writes independent value positions together, then handles commands and answers separately, can approach the distribution of fresh training samples. Confidence ranking does not discover that order in their experiment; it writes the values last, after other tokens have introduced dependencies.
The result is bounded to a synthetic task
The finding is not a measurement of a commercial language model or a natural text corpus. The authors use a trained research model on ScanAndAdd, whose distribution and constraints are deliberately simple enough to calculate exactly. They explicitly say they do not establish the size of the distribution gap on any natural corpus, and that their experiments do not reproduce the architectures or decoding strategies of product models.
The paper’s broader claim is narrower than a verdict on diffusion as a whole: when a sampler independently writes dependent positions in parallel, the step cannot reproduce the true joint distribution. Confidence scores may still guide useful sampling, but the experiment shows why strong per-sample results alone cannot establish that a generator matches the distribution it was trained to model.