跳转至

04 Tree、多候选与 verification

用树和多候选扩大 target support 覆盖,同时维护严格 residual 账本;也系统审视 block verification 与 lossy verifier 的收益边界。

9 篇核心论文 167 页核读 更新至 2026

读完这一类,应能回答

  • 多候选为何不能各自独立接受后任取一个成功项?
  • 固定 verification 节点预算下,怎样选择最有价值的前缀闭包树?
  • 近似接受应如何量化 regret、轨迹偏离和序列级质量风险?

推荐阅读路线

  1. SpecInfer: Accelerating Large Language Model Serving with Tree-based Speculative Inference and Verification — ASPLOS 2024,2023。
  2. SEQUOIA: Scalable and Robust Speculative Decoding — arXiv preprint,2024。
  3. Block Verification Accelerates Speculative Decoding — ICLR 2025,2025。
  4. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes — arXiv preprint,2026。

全部精读

年份 论文 Venue 核读页码
2023 SpecInfer: Accelerating Large Language Model Serving with Tree-based Speculative Inference and Verification ASPLOS 2024 1-18
2024 Multi-Candidate Speculative Decoding arXiv preprint 1-15
2024 SEQUOIA: Scalable and Robust Speculative Decoding arXiv preprint 1-27
2024 SpecExec: Massively Parallel Speculative Decoding for Interactive LLM Inference on Consumer Devices arXiv preprint 1-20
2025 Block Verification Accelerates Speculative Decoding ICLR 2025 1-30
2025 HeteroSpec: Leveraging Contextual Heterogeneity for Efficient Speculative Decoding arXiv preprint 1-17
2026 Approximate Speculative Decoding arXiv preprint 1-8
2026 MARS: Unleashing the Power of Speculative Decoding via Margin-Aware Verification arXiv preprint 1-12
2026 Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes arXiv preprint 1-20