Diffusion language models (diffusion LMs) promise something autoregressive models cannot: unmasking many tokens at once instead of writing strictly left to right. But that parallelism hides a trap. Independently sampling several positions from the same partially masked sequence breaks the dependencies between tokens — the distribution you draw from is no longer the model's distribution. The paper calls this factorization error, and earlier work dubbed it the "curse of parallel decoding." Confidence-based parallel decoders such as Fast-dLLM and Fast-dLLM++, and numerical samplers like tau-leaping, are all trading exactness for speed.

The price of exactness: one sequential evaluation per token

In the continuous-time view, the reverse generation of a diffusion LM is a continuous-time Markov chain (CTMC): each event unmasks one token. This exact sampling introduces no discretization or factorization error, but it costs one sequential model evaluation per token — the next evaluation cannot start before the previous one finishes. Making it fast without changing the flavor is the real problem.

SpecCTMC: speculative decoding, ported to the CTMC

arXiv 2611.08362 introduces SpecCTMC, in the spirit of speculative decoding for autoregressive LMs (Leviathan/Chen, 2023), with two key twists. First, it is training-free and needs no separate draft model — drafts reuse predictions the target model has already computed. Second, it evaluates several candidate intermediate states in parallel, then applies an exact acceptance-and-correction procedure: at the first rejected position it resamples from the positive residual, discards all later tentative marks, and restarts drafting from the corrected state. The whole construction rests on an elegant observation: for absorbing-mask generators, event clocks depend only on the number of remaining masks, not on which tokens have been revealed, so the clocks can be sampled in advance and preserved through rejection and correction. Theorem 1 proves that the joint distribution of SpecCTMC's trajectory — event times, generation order, and token values — is identical to direct token-by-token simulation: no time-discretization error, no extra factorization error, and the same pass@k for every k.

The numbers: 1.73-4.38x fewer evaluations, up to 2.88x faster

Across LLaDA-8B-Instruct and Dream-v0-Instruct-7B on five benchmarks (GSM8K, MATH, MBPP, HumanEval, IFEval), SpecCTMC cuts sequential target-model evaluations by 1.73-4.38x. At temperature 1, average accuracy moves from 16.0 to 16.5 and from 14.8 to 14.6 on the two models — essentially at parity. On an NVIDIA B200, measured wall-clock speedups average 1.36-1.55x with a peak of 2.88x on IFEval; on the multimodal LLaDA-V the gains grow to 5.93-6.69x fewer evaluations on MathVerse and MathVista. On GSM8K with K=8 attempts, SpecCTMC reaches the same 71.88% Pass@8 as 256-step tau-leaping using 724 versus 2,056 median cumulative evaluations — roughly 2.84x cheaper at equal coverage. The code is open-sourced under Apache 2.0.

Commentary: honest ceilings and the real prize

A dose of cold water: wall-clock gains trail the reduction in evaluations, because parallel batching is not free — long-prompt GSM8K on Dream is at parity or slower, while short-sequence, branch-heavy HumanEval and IFEval are the big winners. But the direction matters. Greedy decoding always yields a single answer, while repeated stochastic sampling can cover solutions greedy misses; that path used to be locked behind "exact = sequential," and SpecCTMC pushes exact stochastic sampling into an affordable range. Speedups should not silently swap out a model's outputs — fast and faithful is the ticket for diffusion LMs to reach production.