arXiv cs.AI· Zheng Wang, Davis Wertheimer, Yu Chin Fabian Lim, Mudhakar Srivatsa, Raghu K. Ganti, Minjia Zhang, Naigang Wang·· 14 小时前AI 评分29
xPress：为投机解码中的扩散草稿模型补回因果性的并行精炼方法
XPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding
AI 导读
xPress 是一种轻量级因果精炼器，用于修复 dFlash 等块扩散草稿模型因逐位置独立采样而缺失的因果依赖，通过并行精炼一次性协调整个扩散块，无需逐 token 循环。在 Qwen3-8B 上，xPress 于七个数学、代码和对话基准上平均提升接受长度约 30%（最高 +56%），端到端解码吞吐平均提升约 1.3 倍（最高 1.7 倍）。
正文 · 原文
Abstract:Block-diffusion drafters like dFlash generate an entire block of draft tokens in a single forward pass, drastically reducing the overhead of multiple-token drafting in speculative decoding. The crucial final step of the single-pass discrete denoising process involves using the logit distribution at each position to sample conditionally independent tokens. The resulting draft is thus a set of per-position marginals, rather than a joint distribution: no draft token is guaranteed to depend on its predecessors. Such independently sampled marginals tend to produce sequences with tokens that are individually likely, but jointly improbable under the target model's distribution, which verifies each token conditionally. This can cause early rejection and limits acceptance length. To address this, we propose xPress as a means to restore the missing causality in diffusion drafters. xPress is a lightweight causal refiner that reconciles the whole diffusion block at once through parallel refinement, restoring and propagating causal dependencies across the draft without a token-by-token loop. On Qwen3-8B, across seven math, code, and chat benchmarks, xPress raises acceptance length by about 30% on average (up to +56%) and its end-to-end decoding throughput by about 1.3 on average (up to 1.7) compared to the original dFlash diffusion drafter.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.02438 [cs.AI] |
| (or arXiv:2608.02438v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.02438 arXiv-issued DOI via DataCite |
Submission history
From: Zheng Wang [view email]
[v1]
Mon, 3 Aug 2026 16:18:22 UTC (103 KB)
[v2]
Mon, 26 Oct 2026 00:18:30 UTC (161 KB)
来源：arXiv cs.AI · arxiv.org