---
title: "EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is arXiv non-exclusive distribution license (verified on arXiv abs page 2026-11-12); this pointer note is our own words and is CC BY 4.0; we link to the paper rather than reproduce it'
source: https://iinventedlight.com/library/notes/2401-15077-eagle.md
source_license: "arXiv non-exclusive distribution license"
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Yuhui Li, Fangyun Wei, Chao Zhang, Hongyang Zhang, EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-12.'
category: arxiv
topic_area: 'ai-research: efficiency'
tags: arxiv, efficiency, speculative-decoding, inference, eagle
arxiv_id: 2401.15077
alias: 2401-15077-eagle
type: pointer
format: pointer
sources:
  - id: S1
    title: "EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty"
    publisher: arXiv
    authors: "Yuhui Li, Fangyun Wei, Chao Zhang, Hongyang Zhang"
    url: https://arxiv.org/abs/2401.15077
    version_read: "v3 (4 Mar 2025)"
    accessed: 2026-11-12
    license: "arXiv non-exclusive distribution license"
---
# EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty

## Summary
[EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty](https://arxiv.org/abs/2401.15077) (2024) is a speculative-decoding method that drafts from second-to-top-layer features rather than from tokens alone, and then accounts for the uncertainty in that feature-level prediction.

What it introduced: the Extrapolation Algorithm for Greater Language-model Efficiency (EAGLE). The authors argue that autoregression is easier at the feature level than at the token level, but that leftover uncertainty at that level limits quality unless it is handled. Their fix is to feed a token sequence advanced by one step so the feature predictor can be accurate with little extra cost. On LLaMA2-Chat 70B they report a 2.7x to 3.5x latency speedup and doubled throughput, without changing the output distribution. Tests also cover Vicuna, Mixtral, and tasks in dialogue, code, math, and instruction following.

What it led to: a widely deployed speculative-decoding method in serving stacks. Live notes: [2211-17192-speculative-decoding](https://iinventedlight.com/library/notes/2211-17192-speculative-decoding/) and [ao-cost-speculative-decoding](https://iinventedlight.com/library/notes/ao-cost-speculative-decoding/).

Where to read it: https://arxiv.org/abs/2401.15077 (v3, last revised 4 Mar 2025).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
