Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

  • 类型:arxiv
  • 标识:2607.26627
  • 链接:https://arxiv.org/abs/2607.26627
  • 主分类:llm-infra
  • 形态:method
  • 被引:0
  • 被引来源:Semantic Scholar + OpenAlex
  • S2被引:0
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:This work presents a principled analysis of the distributions induced by lossy verification methods, and identifies a fundamental pitfall: performance can degrade significantly compared to the true truncation sampling baseline due to distributional distortion.
  • OpenAlex ID:W7171842053
  • OpenAlex DOI:10.48550/arxiv.2607.26627
  • DOI:10.48550/arxiv.2607.26627
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2607.26627
  • OpenAlex更新:2026-08-26
  • 待LLM分类:否
  • 标题中文:重思 Speculative Decoding 中的有损验证:机制、权衡与失败模式
  • TLDR中文:对有损验证方法所诱导的分布进行原则化分析,揭示一项根本缺陷:由于分布失真,其性能可能相对于真实截断采样基线显著退化。
  • 来源文件
  • /inbox/tom/_candidates/2026-07-31-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]