Parallel drafts stop guessing alone: candidate rescoring lifts speculative decoding throughput and still beats the uncorrected drafter at 32 concurrent streams

The errors in a parallel draft block are selection errors rather than missing candidates: the drafter usually already holds the token the verifier eventually commits among its leading candidates, it simply does not pick it. Autoregressive drafters generate position by position conditioned on the realized prefix, so intra-block dependencies come for free; a parallel diffusion drafter predicts the whole block in one pass and each position cannot see the token identities actually selected at preceding positions, so individually plausible predictions can form an incoherent continuation. Recording DFlash candidates during speculative decoding on Qwen3-8B and comparing them with the realized continuation, top-16 coverage of the verifier's token is about 0.951 at the first drafted slot and about 0.608 at the last, while top-1 falls to about 0.154; the top-16 to top-1 gap widens from first to last slot on all four verifiers.

Correction is cast as rescoring and re-selecting the drafter's existing candidates rather than generating new tokens, so no additional drafter forward pass is needed. Earlier correction methods DSpark and Domino both introduce intra-block dependencies position by position, placing the added module on the critical path of every drafting cycle, so its latency cost grows with load. The corrector uses only the hidden states, candidate token ids and logits left by a single DFlash pass, scores transitions between candidates at adjacent slots in one batched operation, then walks the grid left to right with a prefix-first greedy rule; on Qwen3-8B the corrector has 2,229,505 parameters and is trained jointly with the drafter.

The throughput gain holds and widens under concurrent serving because the per-cycle cost of parallel correction does not grow with load. Sequential correctors progressively lose their advantage as concurrency rises, and Domino and DSpark become slower than the uncorrected DFlash drafter at the highest concurrency. On Qwen3-8B with thinking off, from 4 to 32 concurrent streams DCOCO has the highest throughput at every concurrency, leading the strongest baseline by 6 to 23 percent under greedy decoding; across the ten benchmarks it averages 959.7 tokens per second under greedy decoding and 762.2 under sampling, above DFlash, Domino and DSpark.

The report covers four verifier backbones (Qwen3-4B, Qwen3-8B, Qwen3-Coder-30B-A3B, Llama-3.1-8B) and ten benchmarks spanning mathematics, code and general-purpose generation, served with vLLM on a single B200 GPU, with block size 16 and at most 2048 generated tokens without thinking and 4096 with extended thinking. Throughput at concurrency 1 is the median request-level rate, sampling results are reported as mean and standard deviation over seeds 42, 43 and 44, and the concurrency study spans 4 to 32 streams. A reader interested in sequential correctors at low concurrency, or in behaviour on parallel drafters other than DFlash, would need to check elsewhere.

The corrector's gain depends on the drafter's candidate pool actually containing the token the verifier commits, and that coverage is about 0.608 at the last drafted slot, so not every position has a correct candidate to re-select. Training labels come from the verifier's argmax under a teacher-forced context, and a transition is supervised only when both slots' verifier tokens lie inside their top candidate sets. Each method is trained once with a single seed, so variability across training seeds is not measured; the reported sampling standard deviations quantify evaluation variability at fixed checkpoints. On path selection, prefix-first greedy achieves higher throughput than geometrically discounted global selection on all nine evaluated suites, while their accepted lengths are close, averaging 5.254 and 5.258.