A new decoding strategy uses the model itself to measure how committing one token changes the confidence of another – a quantity called coupling. By probing just the local block of the current forward pass, the method avoids a full extra pass for every candidate token. The approach keeps the same mean‑field energy rule but swaps the similarity penalty for this coupling.

Across three language models and four datasets, the method requires far fewer decoding steps than the previous mean‑field baseline and is faster in most settings, even when the key‑value cache is used. The average accuracy loss is only 1.1 percentage points. Shuffling the couplings or using a uniform bonus removes this advantage, confirming that the measured interaction between tokens is crucial.

Each scoring step with block‑local probes costs roughly 1.2 to 1.5 times a plain mean‑field step, so the wall‑clock speed‑up is somewhat lower than the step‑count gain. Still, the method achieves 26 % to 55 % fewer steps, making the extra per‑step cost offset by the reduced overall iteration count.

The technique relies on no extra training or model – the original diffusion model provides the necessary predictions. It simply recomputes the block for each candidate token and reads the changed confidence at other positions, a cheap operation validated by a Pearson correlation of 0.96 with a full forward‑pass probe.

Verified fact evidence: measured coupling, mean‑field comparison, diffusion model capability, and step‑cost limitation.