Speculative decoding speeds up large language model inference by having a lightweight drafter propose several tokens at once, which the target model then verifies in parallel. Diffusion-based drafters have made this approach especially appealing, since deep attention-based backbones produce strong predictions while drafting many future positions simultaneously instead of one at a time.
That parallelism has a drawback. Tokens inside a draft block are predicted without knowing the realized identities of the tokens that precede them in the same block, so the intra-block dependencies that autoregressive generation would naturally provide are missing.
To address this, researchers introduce DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores those dependencies while keeping drafting parallel. Rather than committing to each position's top prediction, DCOCO keeps multiple candidate tokens and evaluates their contextual compatibility across adjacent draft positions in parallel, allowing contextually informed selection at low additional cost.
According to the paper, DCOCO achieves the highest average throughput in every evaluated configuration across four verifier backbones and ten benchmarks, covering greedy and stochastic decoding with and without extended reasoning, with speedups of up to 6.6x over vanilla decoding.
The advantage also holds under concurrent serving: DCOCO reportedly outperforms the strongest competing method by 6–23% across 4–32 concurrent requests, while other sequential correctors fall below uncorrected DFlash at those loads. The authors conclude that recovering intra-block dependencies through parallel contextual scoring yields consistent acceleration across decoding regimes and serving loads.