DCOCO Hits 6.6x LLM Decoding Speedup, Beats Rivals Under Load

A new 'contextual corrector' for diffusion-based speculative decoding reaches up to 6.6x speedup over vanilla LLM decoding and holds a 6-23% throughput lead over the strongest rival across 4-32 concurrent requests, according to a preprint on arXiv.

The method, called DCOCO (DFlash Contextual Corrector), targets a specific weakness in diffusion drafters: tokens inside a draft block are predicted in parallel, so each slot is picked without seeing "the realized identities of preceding draft tokens," the authors write. Rather than commit to the top prediction at every position, DCOCO keeps several candidates per slot and scores their "contextual compatibility across adjacent draft positions in parallel."

Across four verifier backbones and ten benchmarks, the paper reports "the highest average throughput in every evaluated configuration," spanning greedy and stochastic decoding with and without extended reasoning. The production-relevant finding sits in the concurrent-serving numbers: at 4-32 simultaneous requests, "other sequential correctors fall below uncorrected DFlash at these loads." In other words, the rival class of fix actively costs throughput once the server is busy.

The abstract gives a 6-23% range rather than per-configuration numbers and publishes no wall-clock figures, hardware details, or per-benchmark breakdown. It is a preprint, not peer-reviewed.