A new arXiv preprint, "Factual Access States in Diffusion Language Models: From Mechanism to Knowledge Editing and Retrieval" (2610.27580), argues that factual recall in diffusion language models behaves like a lookup. A representation inside the network serves as the key, and the properties of that key shape which queries can access the fact.
The authors suggest this framing has direct consequences: an edit must cover whatever the key covers, and a retriever can only read what the key makes available. The abstract is brief and does not yet spell out the proposed mechanism in full, but it positions "factual access states" as the central object connecting model internals to editing and retrieval.
Because there is only one source, there are no conflicting accounts to compare; the summary reflects the preprint's own framing.