06 Serving、基准、安全与应用

把 speculative decoding 放进真实系统：dynamic batching、goodput、SLA、MoE、RL rollout、统一基准，以及输出不变但计算成本崩溃的攻击。

11 篇核心论文 190 页核读 更新至 2026

读完这一类，应能回答

• 为什么 batch=1 的最高 speedup 不能代表生产 serving？
• 怎样用 non-anticipating scheduler 给全 batch 分配 draft / verify 预算？
• 如何隔离 acceptance-collapse 攻击并给出最坏额外成本上界？

推荐阅读路线

1. The Synergy of Speculative Decoding and Batching in Serving Large Language Models — arXiv preprint，2023。
2. Speculative Decoding: Performance or Illusion? — MLSys 2026，2026。
3. SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding — ICML 2026，2026。
4. Adversarial Prompts for Acceptance Collapse in Speculative Decoding — arXiv preprint，2026。

全部精读

年份	论文	Venue	核读页码
2023	The Synergy of Speculative Decoding and Batching in Serving Large Language Models	arXiv preprint	1-9
2025	Speculative Streaming: Efficient and Scalable Speculative Decoding with Multi-Stream Attention	EMNLP 2025	1-24
2026	Accelerating Large-Scale Reasoning Model Inference: Self-Speculative Decoding with Sparse Attention (SparseSpec)	MLSys 2026	1-15
2026	AcceptMoE: Commitment-Weighted Self-Sizing Verifier Expert Sets for Efficient MoE Speculative Decoding	arXiv preprint	1-10
2026	Adversarial Prompts for Acceptance Collapse in Speculative Decoding	arXiv preprint	1-15
2026	Lossless but Not Free: An Empirical Anatomy of Speculative Decoding on Consumer Hardware	arXiv preprint	1-15
2026	Mistletoe: Stealthy Acceleration-Collapse Attacks on Speculative Decoding	arXiv preprint	1-14
2026	PRISM: Parametrically Refactor Inference for Speculative Decoding Draft Models	MLSys 2026	1-14
2026	SpecRoll: Fast-Slow Verifier-Feedback Adaptation for Speculative Reinforcement Learning Rollouts	arXiv preprint	1-23
2026	Speculative Decoding: Performance or Illusion?	MLSys 2026	1-23
2026	SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding	ICML 2026	1-28

继续横向比较

读完单篇后回到方法矩阵，检查正确性保证、proposal 结构和系统工作区间是否可直接比较。

打开跨论文比较 返回全部目录