---
title: 'Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision'
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 2026-11-11); 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/2312-09390-weak-to-strong-generalization.md
source_license: 'arXiv non-exclusive distribution license'
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Collin Burns and 11 coauthors (listed on the arXiv abs page), Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-11
status: published
approved_by: "Cole Withington, 2026-11-11 20:58 PT, Grok Bot chat widget t1717s7407: \"Publish all 100 including the LLMs Encode Clinical Knowledge note\""
checked_on: 2026-11-11
as_of: 2026-11-11
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-11.'
category: arxiv
topic_area: 'ai-research: alignment-and-safety'
tags: arxiv, alignment, supervision, evals
arxiv_id: 2312.09390
alias: 2312.09390-weak-to-strong-generalization
type: pointer
format: pointer
sources:
  - id: S1
    title: 'Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision'
    publisher: arXiv
    authors: 'Collin Burns and 11 coauthors (listed on the arXiv abs page)'
    url: https://arxiv.org/abs/2312.09390
    version_read: 'v1 (14 Dec 2023)'
    accessed: 2026-11-11
    license: 'arXiv non-exclusive distribution license'
---
# Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

## Summary
[Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision](https://arxiv.org/abs/2312.09390) (2023) is OpenAI's study of a basic oversight question: if the supervisor is weaker than the model, can that supervisor still elicit the stronger model's capabilities reliably? That question matters once models start to outpace human evaluators on some tasks.

What it introduced: a set of experiments in which a weaker model (or weaker signal) is used to supervise a stronger one, and measurements of when the strong model still generalizes in the intended direction.

What it led to: later work on scalable oversight, including debate and preference-training debates about who is actually in control of the reward. Related live notes: [2204-05862-helpful-harmless-rlhf](https://iinventedlight.com/library/notes/2204-05862-helpful-harmless-rlhf/). This batch also points to AI safety via debate.

Where to read it: https://arxiv.org/abs/2312.09390 (v1, 14 Dec 2023).

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).