Found first: a primary source the press has not covered yet.
In diffusion language models, wrong answers on math problems require more forward passes per token than correct ones. Mohsen Fayyaz, Ali Modarressi, Daniel Israel, Lucas Bandarkar, and Nanyun Peng show in a new paper that weighting candidates by decoding speed beats majority voting by up to 4.2 points at no extra compute cost. Their study covers two diffusion LMs across 10 tasks.
What the source says
The authors analyze forward-pass counts across 10 tasks and find the cost varies sharply with output type. In DiffusionGemma, story writing requires more than three times as many forward passes as grammar correction. The driving factors are output novelty relative to the input and word rarity in context. Highly tokenized languages compound cost further, adding to their existing token-count disadvantage. On math problems, the same property becomes a signal: correct answers converge faster, wrong answers do not, and ranking candidates by decoding speed beats majority voting by up to 4.2 points at no added cost.
Why it matters
Majority voting is a standard reliability technique for language model inference: generate several candidates and pick the most frequent answer. It works, but it multiplies compute by the number of candidates. If decoding speed already separates right from wrong, practitioners using diffusion LMs can rank outputs without the extra runs. The finding is specific to diffusion LMs, which refine multiple tokens across successive forward passes, making pass count a natural measurable rather than an approximation. The task-cost variation has separate implications for capacity planning: open-ended writing burns more than three times the compute of constrained editing at the same output length.