Limits of Confidence in Diffusion | AI Research

Limits of Confidence in Diffusion
This paper investigates the fundamental limitations of discrete diffusion models—such as remasking and uniform-state samplers—which generate sequences by predicting multiple token positions simultaneously. While these models are designed to produce samples efficiently by writing several tokens at once, the authors demonstrate that this parallel approach often fails to match the true distribution of the training data. The core issue arises because these models rely on per-position probability distributions, which cannot capture the complex dependencies between tokens that exist in domains like language or image generation. The same large language models question is explored in LimiX-2, which adds a research perspective.

The Problem with Parallel Generation

Discrete diffusion models typically generate sequences by choosing a group of positions to fill in a single step. The authors show that a step can only accurately reflect the training distribution if the chosen positions are conditionally independent given the tokens already fixed. When a model writes multiple dependent positions simultaneously, it introduces a "distribution shift." This shift occurs because the model treats these positions as independent, effectively ignoring the underlying constraints of the data. Even if a model is perfectly trained to predict individual tokens, it will still produce the wrong joint distribution if it attempts to write dependent tokens in parallel.

Why Per-Position Distributions Fall Short

A key finding is that per-position distributions are insufficient to determine whether a group of tokens is independent. Two different datasets can have identical individual token probabilities but completely different rules for which combinations of tokens are allowed to appear together. Because the model only "sees" individual probabilities, it has no way to certify that a group of positions is safe to write simultaneously. Consequently, the model may inadvertently assign probability to impossible sequences—states that do not exist in the training data—because it cannot account for the "entangled" nature of the tokens it is generating. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.

Insights from the ScanAndAdd Task

To verify these theoretical limits, the authors used a synthetic task called ScanAndAdd, where the true joint distribution is known in closed form. They found that even when a model achieves perfect "correctness" and "well-formedness" on individual metrics, it remains significantly far from the true data distribution. Specifically, when the model used a confidence-based ranking to choose which positions to write, the generated distribution was 29 times further from the target than the sampling-noise floor. In contrast, when the model was forced to follow a hand-specified, sequential write order, it produced a distribution that was indistinguishable from the training data.

Implications for Model Design

The research suggests that the common practice of using confidence rankings to determine write order does not solve the underlying dependency problem. Because these models cannot verify the independence of the groups they write, they are prone to errors that are not captured by standard per-sample metrics. The authors conclude that for a sampler to be truly accurate, it must avoid writing dependent groups in parallel. However, since the model cannot identify these dependencies on its own, this creates a fundamental trade-off between the speed of parallel generation and the accuracy of the resulting distribution. The ai search story also surfaces in New AI Architecture Mimics the Human..., adding another angle. as detailed in the full paper on Arxiv