Skip to content

Paper 039 / Anthropic

Language Models (Mostly) Know What They Know

The paper evaluates whether language models can estimate the probability that their own answers are correct, including via verbalized confidence and learned probes.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / evaluation

The paper evaluates whether language models can estimate the probability that their own answers are correct, including via verbalized confidence and learned probes.

CONTRIBUTION / 01

Contribution 1

Measured self-evaluation across knowledge and multiple-choice tasks.

CONTRIBUTION / 02

Contribution 2

Compared direct probability predictions with verbal confidence formats.

CONTRIBUTION / 03

Contribution 3

Studied calibration, task transfer, and the relationship between confidence and correctness.

02

Research context

evaluation / 2022

Language Models (Mostly) Know What They Know places calibration inside the broader evaluation discussion at Anthropic, with uncertainty supplying a second analytical lens. The editorial sequence connects three claims: Measured self-evaluation across knowledge and multiple-choice tasks; Compared direct probability predictions with verbal confidence formats; and Studied calibration, task transfer, and the relationship between confidence and correctness. The combination matters because self evaluation only has meaning under the paper's stated setup. Operationally, the record points to one consequence: model confidence can be informative, but it must be calibrated on the relevant model and task before being shown as reliability.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Language Models (Mostly) Know What They Know tests measured self-evaluation across knowledge and multiple-choice tasks and compared direct probability predictions with verbal confidence formats against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

For Language Models (Mostly) Know What They Know, the useful question is what the documented setup supports and where transfer remains untested. Its evidence chain begins with two documented moves: Measured self-evaluation across knowledge and multiple-choice tasks; and Compared direct probability predictions with verbal confidence formats. Reported evidence then addresses: Studied calibration, task transfer, and the relationship between confidence and correctness. The resulting interpretation is practical but conditional: model confidence can be informative, but it must be calibrated on the relevant model and task before being shown as reliability. Its boundary is that reading Language Models (Mostly) Know What They Know as evidence for calibration requires preserving selected threat model, evaluator models, prompt sampling, model revisions, construct validity, and rater instructions; changing those conditions creates a new experiment. Any extension should report how altered uncertainty conditions affect the original calibration result. An independent check of Language Models (Mostly) Know What They Know needs a fixed calibration comparison, a declared uncertainty variation, and saved cases where Compared direct probability predictions with verbal confidence formats does not carry over.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Language Models (Mostly) Know What They Know supports studied calibration, task transfer, and the relationship between confidence and correctness.
05

Practical implication for AI builders

Anthropic / 2022

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / calibration

For a proposed BrokenGPT experiment based on Language Models (Mostly) Know What They Know, calibrate confidence per endpoint and task family, and use it to trigger retrieval or human escalation rather than presenting raw self-reports as certainty. Keep the calibration path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / uncertainty

Validate the proposed calibration route against the paper's reported outcome: Studied calibration, task transfer, and the relationship between confidence and correctness. Before a proposed Language Models (Mostly) Know What They Know change advances, compare calibration, slice stability, metric coverage, and evaluator agreement and inspect uncertainty counterexamples outside the aggregate calibration result.

TRADEOFF / 03

Proposed decision boundary / self evaluation

Balance breadth, repeatability, and construct validity before promoting the proposed self evaluation design. Because reusing the mechanism calls for separate evidence about deployment drift, language coverage, unsampled behaviors, adversarial adaptation, and judge bias, not an inference from the original benchmark alone, 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

  • Reading Language Models (Mostly) Know What They Know as evidence for calibration requires preserving selected threat model, evaluator models, prompt sampling, model revisions, construct validity, and rater instructions; changing those conditions creates a new experiment.
  • Reusing the mechanism calls for separate evidence about deployment drift, language coverage, unsampled behaviors, adversarial adaptation, and judge bias, not an inference from the original benchmark alone.

PRIMARY SOURCES

  1. 01
    Language Models (Mostly) Know What They Know

    Anthropic — Primary primary arXiv paper / 11 July 2022 / Saurav Kadavath, Tom Conerly, Amanda Askell, and 33 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Language Models (Mostly) Know What They Know study?

The paper evaluates whether language models can estimate the probability that their own answers are correct, including via verbalized confidence and learned probes.

02Which methods does Language Models (Mostly) Know What They Know use?

The experimental design in Language Models (Mostly) Know What They Know tests measured self-evaluation across knowledge and multiple-choice tasks and compared direct probability predictions with verbal confidence formats against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Language Models (Mostly) Know What They Know report?

The reported evidence in Language Models (Mostly) Know What They Know supports studied calibration, task transfer, and the relationship between confidence and correctness.

04What is the proposed BrokenGPT application for Language Models (Mostly) Know What They Know?

Proposed: calibrate confidence per endpoint and task family, and use it to trigger retrieval or human escalation rather than presenting raw self-reports as certainty.

MAJOR LAB RESEARCH / PAPER 039

Continue after Language Models (Mostly) Know What They Know

After Language Models (Mostly) Know What They Know, browse the full index for adjacent evaluation research and work from Anthropic.

Open the paper index