Contextual Correction of Parallel Draftsfor Speculative Decoding
DCOCO tries to recover the token-to-token context that parallel speculative drafts lose.
The paper introduces a lightweight corrector that keeps multiple candidate tokens per draft position and scores how well adjacent candidates fit together. It is meant to preserve parallel drafting while avoiding the blind commitment to each position’s top guess. Across four verifier backbones and ten benchmarks, the authors report the highest average throughput in every tested setup, with speedups up to 6.6x over vanilla decoding. Under concurrent serving, they say DCOCO beats the strongest competing method by 6-23% across 4-32 simultaneous requests. Source: HF Daily Papers' note.
The paper introduces a lightweight corrector that keeps multiple candidate tokens per draft position and scores how well adjacent candidates fit together. It is meant to preserve parallel drafting while avoiding the blind commitment to each position’s top guess. Across four verifier backbones and ten benchmarks, the authors report the highest average throughput in every tested setup, with speedups up to 6.6x over vanilla decoding. Under concurrent serving, they say DCOCO beats the strongest competing method by 6-23% across 4-32 simultaneous requests. Source: HF Daily Papers' note.
score 4