SPOT: commit-order prediction enables 3.2x speculative decoding speedups for diffusion language models

By training a 0.62M-parameter drafter to predict position commit trajectories and introducing coupled-noise match verification, SPOT accelerates diffusion LLMs without quality loss.

Primary source: arXiv:2610.22543: https://arxiv.org/abs/2610.22543

What’s new: Diffusion language models (DLMs) offer an attractive alternative to autoregressive architectures by allowing non-sequential, parallel token generation across masked positions. However, exact generation in DLMs remains slow because standard unmasking decodes only one position per full model forward pass, while naive multi-token unmasking introduces substantial generation artifacts. Speculative decoding provides lossless acceleration for autoregressive models, but directly applying it to diffusion models is non-trivial because DLMs must determine both which token to generate and which position to commit next. In On the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction (arXiv:2610.22543, 15 Oct 2026), researchers Eliezer Shlomi, Eliron Rahimi, Avi Mendelson, and Chaim Baskin present Speculating Positions Order Trajectory (SPOT), an architecture that unlocks lossless speculative acceleration for diffusion language models.

The authors observe an empirical property in DLM generation: a model’s intermediate token predictions from the previous pass already correctly identify the target token at 92% to 100% of candidate positions, meaning the primary bottleneck is predicting the optimal unmasking order. SPOT solves this by introducing an ultra-lightweight 0.62M-parameter drafting model that inspects the target model’s output distributions and predicts a tree of candidate commit orders. The full target model then evaluates the proposed candidate tree in a single forward pass by verifying state consistency. To overcome the limitation of existing verification methods that assume zero temperature ($T=0$) or sequential proposals, the authors develop coupled-noise match verification, utilizing shared-Gumbel-noise coupling to validate diffusion blocks at arbitrary temperatures. In extensive benchmarks across six diverse DLMs, SPOT reduced target model forward passes by up to 6.8x and delivered up to a 3.2x throughput improvement over standard decoding, consistently outperforming calibrated and training-free drafting baselines at both $T=0$ and $T=1$. Furthermore, the compact drafter demonstrated strong zero-shot transferability across different base DLMs without retraining.

Why it matters: Slow generation speeds have been a primary impediment to the practical adoption of diffusion language models in production environments. By demonstrating that position commit order can be learned with minimal compute overhead and verified in parallel across arbitrary sampling temperatures, SPOT removes a major performance barrier and makes diffusion-based text generation competitive with optimized autoregressive serving engines.

Caveats: While the 0.62M-parameter drafting network adds negligible parameter overhead, managing speculative tree verification increases peak memory consumption and complicates KV-cache and batch scheduling in memory-constrained environments. Additionally, on highly ambiguous or open-ended creative generation tasks where token prediction entropy is high, the acceptance rate of multi-token speculative branches may diminish.