---
title: 'Learning to summarize from human feedback'
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/2009-01325-learning-to-summarize.md
source_license: 'arXiv non-exclusive distribution license'
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Nisan Stiennon, Long Ouyang, Jeff Wu, and 6 coauthors (listed on the arXiv abs page), Learning to summarize from human feedback (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: alignment-and-safety'
tags: arxiv, alignment, rlhf, summarization
arxiv_id: 2009.01325
alias: 2009.01325-learning-to-summarize
type: pointer
format: pointer
sources:
  - id: S1
    title: 'Learning to summarize from human feedback'
    publisher: arXiv
    authors: 'Nisan Stiennon, Long Ouyang, Jeff Wu, and 6 coauthors (listed on the arXiv abs page)'
    url: https://arxiv.org/abs/2009.01325
    version_read: 'v3 (15 Feb 2022)'
    accessed: 2026-11-12
    license: 'arXiv non-exclusive distribution license'
---
# Learning to summarize from human feedback

## Summary
[Learning to summarize from human feedback](https://arxiv.org/abs/2009.01325) (2020) showed that a summarization policy can be improved by learning a reward model from human comparisons and then optimizing the policy against that reward, rather than relying only on reference summaries and ROUGE. It is the direct language-task precursor of later assistant RLHF.

What it introduced: preference-based reward learning and reinforcement learning for summarization, with human comparisons as the supervision signal.

What it led to: InstructGPT-style RLHF for chat models, and then alternatives such as direct preference optimization. Live notes: [2203-02155-instructgpt](https://iinventedlight.com/library/notes/2203-02155-instructgpt/), [1706-03741-rl-from-human-preferences](https://iinventedlight.com/library/notes/1706-03741-rl-from-human-preferences/), [1707-06347-ppo](https://iinventedlight.com/library/notes/1707-06347-ppo/), and [2305-18290-direct-preference-optimization](https://iinventedlight.com/library/notes/2305-18290-direct-preference-optimization/).

Use it when you want the summarization paper that sits immediately before InstructGPT, and you need the human-comparison setup rather than a ROUGE-only story. Where to read it: https://arxiv.org/abs/2009.01325 (v3 (15 Feb 2022)).

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).
