Masked diffusion language models (MDLMs) — LLaDA, Dream and relatives — generate text by iteratively unmasking tokens in a partially masked sequence. The pitch is obvious: unlike left-to-right autoregression, you can commit several tokens per step. The catch is that the tokens are not independent. Unmask them independently and you get a different output distribution than the model was trained to produce, with no clean bound on how wrong it got.
Two escape routes have dominated. Approximate solvers (τ-leaping, θ-trapezoidal) freeze transition rates over finite intervals, buying speed and quietly changing the law. Parallel decoders (Fast-dLLM and its descendants) pick high-confidence positions, which is faster still but bakes in a factorization error the paper quantifies: for a setting with δ = 0.05, independent parallel sampling provably needs at least ~0.897n extra target-model evaluations to compensate.
SpecCTMC takes a third route: keep the exact continuous-time Markov chain (CTMC) and make the evaluations parallel rather than the commits. The insight is structural. In absorbing-mask generators, event times depend only on diffusion time and the number of remaining masks — not on the identities of unmasked tokens. So the event clock can be sampled in advance and survives rejection.
How It Works
The sampler drafts successive token updates using previously computed predictions, evaluates those tentative intermediate states in one batched forward pass, then verifies each proposal sequentially with an exact acceptance-and-correction procedure. Rejecting a proposal does not disturb the sampled clock, because a corrected mark is still a valid mark for that event time. The proof carries over from distribution-preserving speculative sampling for autoregressive models, extended to continuous diffusion: the resulting trajectory has the same joint distribution over event times, generation order and token values as direct simulation. No time-discretization error, no additional factorization error.
The Numbers
On LLaDA-8B-Instruct and Dream-v0-Instruct-7B across GSM8K, MATH, MBPP, HumanEval and IFEval, SpecCTMC uses 1.73–4.38× fewer sequential target-model evaluations than direct CTMC simulation, with measured wall-clock speedups up to 2.88× on an NVIDIA B200. On GSM8K with LLaDA at 256 generated tokens, target NFEs drop by 2.88× while expected accuracy is preserved.
The speedup is not uniform across lengths, and the paper is upfront about it: at L=32 the average wall-clock gain on LLaDA is only 1.05× even though NFE drops 2.58×, because batching cannot amortize a short generation. The best gains land on IFEval and HumanEval (up to 2.07× wall-clock, 3.30× NFE average on LLaDA), where instructions and code give the drafter more to work with. Draft depth of 2–7 is used depending on benchmark.
Two details matter more than the headline. Accuracy differences against DirectCTMC are statistically flat — every paired 95% confidence interval contains zero. And on GSM8K, multiple stochastic attempts solve questions that deterministic greedy denoising misses; SpecCTMC reduces the cumulative evaluation cost of that best-of-N strategy, which is the scenario where exact stochastic sampling actually pays.
The honest caveat: this is a sampler-level win, not an architecture-level one. It reduces how many times you call the model, not how expensive the model is. It also compounds with KV caching and other per-evaluation optimizations rather than replacing them.
Paper: arXiv:2611.08362 · code