The core promise of diffusion language models is parallelism: one forward pass can commit several tokens at once, so decoding should cost less than the token-by-token march of autoregressive models. But how much less, task by task, had never been measured systematically. Researchers from UCLA NLP put two diffusion models through ten tasks and published the ledger on October 19 as arXiv:2610.25835. The headline: the savings depend heavily on the task, and on the most expensive task the advantage disappears entirely.

The Ledger: 3.7 Versus 101 Passes

The metric is forward passes per 100 generated tokens. An autoregressive model always sits at 100; diffusion models promise to stay under that line. The two models tested are DiffusionGemma-26B-A4B, a 26B mixture-of-experts model generating 256-token blocks, and DreamReasoner, an 8B dense model generating 32-token blocks. Both use semi-autoregressive block diffusion and commit only tokens that clear a confidence threshold, so step counts adapt on their own.

On the DiffusionGemma side, grammar correction costs 3.7 passes, extractive summarization 4.3, code editing 5.0, translation 6.0. Open-ended work gets expensive fast: WildBench dialogues cost 8.3, abstractive summarization 9.5, and story writing 11.4 — more than three times grammar fixing. DreamReasoner is pricier overall, climbing from 22.2 passes on grammar correction to 101.0 on abstractive summarization, a spread the paper reports as 4.3x. That 101 is a watershed number: it crosses the autoregressive line of 100, meaning diffusion not only stopped saving money on this task but spent more. Yet the two models — different in architecture, size, and block length — agree almost perfectly on which tasks are expensive.

What Flips the Switch: New Information

Why is a story costly and a grammar fix cheap? The authors test two candidate properties: how novel the output is relative to the input, and how rare its words are within the task. The strongest correlate of cost is neither alone but their combination, a "new information" measure: Pearson +0.76 on DiffusionGemma and +0.85 on DreamReasoner, while surprisal alone manages only +0.18 and −0.09. In plain terms, cost is not driven by obscure vocabulary but by how much of the text is being written for the first time.

The anatomy of a single competition math problem makes the point. The model thinks for 10,339 tokens across 40 blocks, each taking 4 to 22 forward passes (cost 3.81), then writes a 1,241-token answer (cost 2.88). The answer is cheap because it mostly rewords what the thinking already contained; when the answer must add new content the thinking never produced — as in WildBench or story writing — its cost overtakes the thinking itself.

Wrong Answers Burn More, and the Burn Is a Signal

Across 160 competition problems from 5 contests, five seeds each, accuracy falls from 79% through 57% to 44% as problems get harder, while median thinking length grows from 4,416 to 9,710 tokens. The key finding: after controlling for problem difficulty and generation length, incorrect answers still consume more forward passes than correct ones, and wrong answers stop to correct themselves — "wait, let me re-check" moments — about five times as often as right ones. Compute usage itself leaks the model's confidence.

That yields the paper's most practical trick. When sampling multiple candidate answers and taking a majority vote, weight each vote by how quickly it was decoded. Fast decodes get heavier votes, and accuracy beats plain majority voting by 4.2 points with two candidates and 2.3 points with five, at zero extra cost. Multilingual generation carries what the paper calls a double tax: translating 200 English sentences into 25 languages across 16 scripts shows that highly tokenized languages pay both with more tokens and with more passes per token.

So What

Three intuitions get rewritten. First, the diffusion cost advantage is not a uniform asset: copy, edit, and extract workloads collect the full dividend, while open-ended generation may collect nothing — check your task mix before picking a decoding paradigm. Second, "price per token" is a misleading unit for diffusion models: two outputs of identical length from the same model can differ 3x in real compute, a bill token-priced API providers will eventually have to face. Third, decoding speed, usually discarded as a byproduct, is a free confidence signal — which for anyone scheduling inference may be worth more than another benchmark point.