03 Feature head、MTP 与并行块草稿

方向最密集的主线：从 Medusa / EAGLE feature heads 到 DFlash 并行块，再到 Domino、DSpark、PCTree 对候选内因果性与生产调度的修复。

20 篇核心论文 336 页核读 更新至 2026

读完这一类，应能回答

• parallel marginals 为什么不是一个合法的 joint proposal？
• 怎样用低成本 Markov / RNN / tree conditioning 修复 block 内依赖？
• proposal 质量、draft latency 与 verification capacity 应如何联合优化？

推荐阅读路线

1. MEDUSA: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads — ICML 2024，2024。
2. EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty — ICML 2024，2024。
3. DFlash: Block Diffusion for Flash Speculative Decoding — ICML 2026，2026。
4. Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding — arXiv preprint，2026。
5. DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation — arXiv preprint，2026。
6. From Chains to Trees: Parent-Conditioned Drafting for Semi-Autoregressive Speculative Decoding — arXiv preprint，2026。

全部精读

年份	论文	Venue	核读页码
2024	EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees	EMNLP 2024	1-12
2024	EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty	ICML 2024	1-14
2024	Hydra: Sequentially-Dependent Draft Heads for Medusa Decoding	COLM 2024	1-17
2024	MEDUSA: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads	ICML 2024	1-27
2024	Recurrent Drafter for Fast Speculative Decoding in Large Language Models	arXiv preprint	1-14
2025	EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test	NeurIPS 2025	1-20
2025	PARD: Accelerating LLM Inference with Low-Cost Parallel Draft Model Adaptation	arXiv preprint	1-18
2026	AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding	arXiv preprint	1-26
2026	CURE: Local Uncertainty Repair for Block-Parallel Speculative Decoding	arXiv preprint	1-9
2026	DBLast: Dependent Block Drafting for Stochastic Speculative Decoding	arXiv preprint	1-12
2026	DeLS-Spec: Decoupled Long-Short Contexts for Parallel Speculative Drafting	arXiv preprint	1-17
2026	DFLARE: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding	arXiv preprint	1-12
2026	DFlash: Block Diffusion for Flash Speculative Decoding	ICML 2026	1-13
2026	Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding	arXiv preprint	1-11
2026	DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation	arXiv preprint	1-33
2026	From Chains to Trees: Parent-Conditioned Drafting for Semi-Autoregressive Speculative Decoding	arXiv preprint	1-18
2026	JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting	arXiv preprint	1-21
2026	P-EAGLE: Parallel-Drafting EAGLE with Scalable Training	arXiv preprint	1-13
2026	TreeFlash: Parallel AR-Approximation for Faster Speculative Decoding	arXiv preprint	1-13
2026	xPress: Parallel Refinement for Diffusion Drafters in Speculative Decoding	arXiv preprint	1-16

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