English

AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters

Machine Learning 2026-07-21 v1 Computation and Language

Abstract

Speculative decoding, in which a lightweight draft model first generates a draft sequence that is then verified in parallel by the target model, has become a prevalent paradigm for accelerating large language model inference. Recent work such as DFlash further boosts drafting efficiency by leveraging diffusion drafters, whose parallel denoising mechanism enables draft generation in a single forward pass. In this work, we uncover a central pitfall of diffusion drafters: bidirectional attention is a double-edged sword. On one hand, it endows the model with parallel generation and global contextual modeling capabilities; on the other hand, this inherent global dependency introduces high variance at both the domain-level and the token-level: acceptance rates fluctuate substantially across different domains, and draft token quality also varies heterogeneously at different token positions. To tackle this issue, we propose AdaFlash framework, comprising two components: (i) an on-policy distillation (OPD) algorithm with reverse-KL divergence tailored for diffusion drafters, bringing stable convergence and effectively reducing domain-level variance; and (ii) an adaptive length head that dynamically adjusts the candidate sequence length on the fly, substantially lowering the verification cost of the target model and handling token-level variance. Experiments demonstrate that AdaFlash consistently improves speedup rate during deployment, with especially significant gains in high-concurrency scenarios, achieving up to approximately 66% higher throughput than previous state-of-the-art methods.

Cite

@article{arxiv.2607.19223,
  title  = {AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters},
  author = {Yu-Yang Qian and Hao-Cong Wu and Chen Chen and Jiacheng Sun and Zhenhua Dong and Peng Zhao and Zhi-Hua Zhou},
  journal= {arXiv preprint arXiv:2607.19223},
  year   = {2026}
}
R2 v1 2026-07-22T20:50:14.223Z