---
id: 20260911-T0-11
title: "Osprey：不绑定目标模型的预训练，让投机解码的起草模型更强"
title_en: "Osprey: Target-Agnostic Pretraining Boosts Speculative Decoding Drafters"
url: https://ai.daily.yangsir.net/daily/20260911-T0-11
issue_date: 2026-09-11
publish_date: 2026-09-10T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2609.09338
---

# Osprey：不绑定目标模型的预训练，让投机解码的起草模型更强

arXiv 论文提出 Osprey，一种面向投机解码的预训练方案。研究指出，投机解码对 LLM 推理加速很关键，但加速效果不稳定：起草模型通常针对单一目标模型的窄分布训练，工作负载一变接受率就崩塌。Osprey 的做法是不绑定特定目标模型进行预训练，从而在分布变化时保持更强的起草能力。若效果成立，可提升推理加速方案在不同任务间的稳定性。

## English Version

**Osprey: Target-Agnostic Pretraining Boosts Speculative Decoding Drafters**

An arXiv paper introduces Osprey, a pretraining approach for speculative decoding. The authors note that while speculative decoding is critical for accelerating LLM inference, the speedup is fragile: drafters are typically trained against a narrow distribution for a single target model, and acceptance rates collapse under workload shifts. Osprey pretrains drafters in a target-agnostic way to stay robust as distributions change. If validated, it could make inference acceleration more stable across tasks.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2609.09338)

**详情页**：https://ai.daily.yangsir.net/daily/20260911-T0-11

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*