简体中文 · English

DBLast: Dependent Block Drafting for Stochastic Speculative Decoding¶

Core mechanism reproduction. The defining operator runs locally. Private production data, large weights and dedicated serving are not reproduced; paper and local results are separate.

Paper information¶

Field	Details
Paper	arXiv 2608.05448
Institution	Huawei Technologies Canada
First public release	2026-08-05 (arXiv v1)
Author code	No official author code in the paper when checked on 2026-08-09.
Adapter	`dblast`
Local reproduction code	`src/auto_research/reproductions/dblast/`

Original paper summary¶

Background and main changes¶

Topic: Speculative decoding. Parallel block drafters often assume positions are conditionally independent and struggle to match the joint distribution at high sampling entropy. DBLast uses a shared low-rank latent mixture to model cross-position dependence and trains for expected verification length.

Architecture¶

flowchart LR
    N0["Target model"] --> N1
    N1["Low-rank dependent drafter"] --> N2
    N2["Block proposal"] --> N3
    N3["Exact verifier"] --> N4
    N4["Accepted prefix"]

Key figure from the paper¶

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equation¶

\(q(y_{1:B}\mid x)=\int p(z)\prod_{b=1}^{B}q(y_b\mid x,z)\,dz\)

Paper results¶

On Qwen3-4B/8B across GSM8K, MT-Bench, HumanEval and creative tasks, accepted draft length improves consistently in high-entropy regimes. The abstract does not give a single aggregate value.

Local reproduction¶

Compare an independent unigram proposal with a rank-4 conditional proposal against a 32-token Markov target. Estimate accepted length by exact speculative acceptance.

auto-research reproduce --paper dblast --dataset-dir data --seed 42

Stable metrics: metrics/public-seed42.json. No checkpoints are committed.

Comparison: Remove the paper-specific mechanism for the baseline while matching split and budget. The experimental arm uses dblast. See public-seed42.json for relative change; percentages across papers are not comparable.

Scope¶

Local results verify execution and direction of comparison, not paper-scale reproduction.
Private features, live traffic and production serving are unavailable; paper online values are references only.
Evolve-compatible architectures are registered as candidates. Serving-only methods stay standalone runnable adapters, not trainable genomes.