Measured Coupling paper — Parallel diffusion decoding by measuring token coupling — What does it mean?

The news. On November 5, 2026, a Tel Aviv University team posted a paper on parallel decoding in diffusion language models. It keeps the mean-field commit rule but swaps its pairwise term for a measured one, and reports fewer decoding steps than mean-field decoding on three models and four benchmarks, at an average accuracy cost of 1.1 points. Read the paper →

Picture a team solving a crossword where every solver fills a different blank at the same moment, and each sees only the squares already inked in. Two answers can each look certain and still clash where they cross. A masked diffusion model is in exactly that position: it starts from a fully masked answer and, in each step, predicts every masked position independently from the same context, so two tokens can each look certain and still form a wrong pair. An autoregressive model avoids this by writing one token per forward pass, at the price of one full pass per token. Diffusion decoding tries to commit several tokens per pass, and the whole game is choosing which ones.

Mean-field decoding makes that choice with a proxy. Each masked position gets a commit score, and a pairwise penalty subtracts from the score when two positions have similar predicted distributions, measured with Jensen-Shannon divergence. In crossword terms, it compares how the two answers' odds look instead of testing what happens when one answer goes in. That misses both directions: two positions with similar distributions may not touch each other, and two with different distributions may depend strongly, like the digits of one number. The paper's own words for the fix: "We replace this proxy with the coupling itself."

The coupling is measured by asking the model. Pencil answer X into its square, re-run the model, and read how the log-margin at position Y moved. That change is the coupling, and its sign matters. A positive value means revealing X makes Y less certain, so committing both together is risky. A negative value means X supports Y: once X is known, Y becomes more certain. Support is what lets a step commit many tokens at once, and a penalty-only proxy has no way to express it. The measured values plug into the same mean-field update, which runs two iterations per step, and positions above the threshold are committed.

Done naively, one probe costs a full forward pass over the sequence, which would erase the gain. The paper makes it cheap in two moves. First, a probe recomputes only the still-masked positions of the current block, and every other position reuses the keys and values from the step's main pass; this is the same KV reuse idea that the Fast-dLLM dual cache applies, here applied to probes. Second, the probes ride inside the next step's main forward pass: normalization, projections, MLP and output head process main tokens and probe tokens together, so the model weights are read once. Only attention is split, because each probe row attends to its own keys at the masked positions and to the main keys and values everywhere else. Since a step's logits are not known before the pass, the probes use the previous step's proposals, and a probe whose token has since changed is dropped. Reading weights once matters because small batches are limited by memory traffic, the same decode bottleneck behind speculative decoding.

Now the numbers on LLaDA-8B-Instruct, GSM8K, from the paper's Table 1. One-token-per-step decoding needs 256 steps for a 256-token answer. Mean-field decoding needs 65.4 steps, a 3.9× step reduction. Measured coupling needs 40.1 steps, a 6.4× reduction, which is 38.7% fewer steps than mean-field (40.1 / 65.4 = 0.613). Steps are not free: the paper states that a step with probes costs 1.2 to 1.5 times a mean-field step. Using that range as an illustration, 40.1 × 1.2 = 48.1 and 40.1 × 1.5 = 60.2 mean-field-step equivalents, against 65.4, so the work falls by roughly 8% to 26% (illustrative arithmetic from the stated range). The measured wall-clock time per answer on this benchmark is 8.10 s → 5.86 s, at an accuracy of 76.57 → 76.27.

Read the result with its limits in view. Averaged over the three models and four benchmarks, accuracy is 1.1 points below mean-field decoding; per benchmark the difference ranges from 6.1 points lower to 1.8 points higher, and the drop is significant on MATH for two of the three models. The wall-clock gain is smaller than the step gain because of the probe cost above. And the authors note that mutual support shows self-consistency, not correctness, so wrong tokens can reinforce each other. The theory is also an interpretation, not a safety proof: it assumes margins add up when tokens are revealed together, and the paper's own check finds the additive prediction and the jointly measured margin on the same side of the threshold for 90% to 93% of remaining positions.

Goes deeper in: LLM Internals → Text Generation → One Token at a Time

Related explainers

• TIPS learned unmask-and-remask decoding — a different answer to the same question: learn the commit choice, and allow hiding tokens again.
• Masked diffusion serving: denoising-step batching — what the serving side does once a diffusion decoder runs many steps per request.
• A diffusion drafter for asynchronous speculative decoding — the reverse pairing: a diffusion model as the drafter inside speculative decoding.