Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Determine whether natural corpora contain sampler-independent conditional dependencies

Determine whether natural language or other natural-data corpora contain conditional dependencies that prevent independent-update diffusion samplers from selecting exact multi-position write groups using their available per-position distributions.

Background

Theorem 2 establishes that per-position distributions cannot certify whether a selected group is conditionally independent, although a particular target distribution may nevertheless permit the groups selected by a sampler. The ScanAndAdd experiment closes this gap for a synthetic distribution by explicitly exhibiting dependent command and answer groups and showing that confidence rankings encounter them.

The authors explicitly leave unresolved whether analogous dependencies, with comparable consequences for exact parallel decoding, occur in any natural corpus. This is distinct from the general theoretical impossibility because it asks for empirical or mathematical characterization on natural data distributions.

References

A given $p$ may still be one on which the groups a sampler happens to write are independent. Section~\ref{sec:instance} closes that gap for ScanAndAdd by exhibiting the dependencies directly, and we do not close it for any natural corpus.

— Limits of Confidence in Diffusion  (2609.20581 - Webb et al., 17 Sep 2026) in Section 5, Limitations

This local predictability diagnostic leaves open whether the document remains coherent, which we examine in Appendix~\ref{app:coherence}.

— Rethinking the Roles of Diffusion and Decoding: Predictive-State Flows for Parallel Language Generation  (2610.12898 - He et al., 9 Oct 2026) in Appendix, Section “Document length, MAUVE and a length-controlled window,” paragraph preceding Table ref{tab:memo}