Skip to content

Paper 036 / OpenAI

Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

Weak-to-strong generalization studies whether a strong model can learn a task from labels produced by a weaker model, using controlled model-size gaps as an analogy for future oversight.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / alignment_and_safety

Weak-to-strong generalization studies whether a strong model can learn a task from labels produced by a weaker model, using controlled model-size gaps as an analogy for future oversight.

CONTRIBUTION / 01

Contribution 1

Formalized a laboratory setting for weak-supervisor versus strong-student training.

CONTRIBUTION / 02

Contribution 2

Measured how much strong capability is recovered under weak labels.

CONTRIBUTION / 03

Contribution 3

Introduced methods and baselines for improving recovery beyond naive imitation.

02

Research context

alignment_and_safety / 2023

Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision places weak to strong inside the broader alignment and safety discussion at OpenAI, with scalable oversight supplying a second analytical lens. The editorial sequence connects three claims: Formalized a laboratory setting for weak-supervisor versus strong-student training; Measured how much strong capability is recovered under weak labels; and Introduced methods and baselines for improving recovery beyond naive imitation. The combination matters because alignment only has meaning under the paper's stated setup. Operationally, the record points to one consequence: oversight quality can bottleneck a model even when the model has greater latent capability than its supervisor.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision tests formalized a laboratory setting for weak-supervisor versus strong-student training and measured how much strong capability is recovered under weak labels against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

For Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision, the useful question is what the documented setup supports and where transfer remains untested. Its evidence chain begins with two documented moves: Formalized a laboratory setting for weak-supervisor versus strong-student training; and Measured how much strong capability is recovered under weak labels. Reported evidence then addresses: Introduced methods and baselines for improving recovery beyond naive imitation. The resulting interpretation is practical but conditional: oversight quality can bottleneck a model even when the model has greater latent capability than its supervisor. Its boundary is that the weak to strong comparison in Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision is interpretable only alongside rater instructions, prompt sampling, model revisions, evaluator models, selected threat model, and construct validity, which limits claims about unseen deployments. Any extension should report how altered scalable oversight conditions affect the original weak to strong result. A credible extension of Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision would freeze its weak to strong reference, perturb scalable oversight deliberately, and publish exceptions to Measured how much strong capability is recovered under weak labels alongside aggregate results.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision supports introduced methods and baselines for improving recovery beyond naive imitation.
05

Practical implication for AI builders

OpenAI / 2023

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / weak to strong

For a proposed BrokenGPT experiment based on Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision, audit evaluator-model limits explicitly and combine weak automated grading with gold human cases and adversarial tests before changing routing policy. Keep the weak to strong path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / scalable oversight

Validate the proposed weak to strong route against the paper's reported outcome: Introduced methods and baselines for improving recovery beyond naive imitation. The proposed Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision test should capture adversarial coverage, helpful-answer retention, and refusal precision, with scalable oversight error slices reported apart from the headline weak to strong result.

TRADEOFF / 03

Proposed decision boundary / alignment

Balance usefulness, oversight burden, and residual risk before promoting the proposed alignment design. Because product evidence would remain incomplete without testing deployment drift, language coverage, unsampled behaviors, adversarial adaptation, and judge bias under the selected scalable oversight workload, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • The weak to strong comparison in Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision is interpretable only alongside rater instructions, prompt sampling, model revisions, evaluator models, selected threat model, and construct validity, which limits claims about unseen deployments.
  • Product evidence would remain incomplete without testing deployment drift, language coverage, unsampled behaviors, adversarial adaptation, and judge bias under the selected scalable oversight workload.

PRIMARY SOURCES

  1. 01
    Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

    OpenAI — Primary primary arXiv paper / 14 December 2023 / Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, and 9 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision study?

Weak-to-strong generalization studies whether a strong model can learn a task from labels produced by a weaker model, using controlled model-size gaps as an analogy for future oversight.

02Which methods does Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision use?

The experimental design in Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision tests formalized a laboratory setting for weak-supervisor versus strong-student training and measured how much strong capability is recovered under weak labels against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision report?

The reported evidence in Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision supports introduced methods and baselines for improving recovery beyond naive imitation.

04What is the proposed BrokenGPT application for Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision?

Proposed: audit evaluator-model limits explicitly and combine weak automated grading with gold human cases and adversarial tests before changing routing policy.

MAJOR LAB RESEARCH / PAPER 036

Continue after Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

After Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision, browse the full index for adjacent alignment_and_safety research and work from OpenAI.

Open the paper index