← Research index
FRONTIER LAB WATCHReviewed 2 September 2026
Self-correction is two problems, not one
Google Research separates finding a mistake from repairing it; GERO turns that distinction into separate evidence contracts.
Xamit KadirbekovIndependent analysis · Source: Google Research
Self-correctionError localizationEvaluation
STATUS · SOURCE REPORT + GERO ANALYSISThis brief has not independently reproduced the laboratory's experiment.
PRIMARY SOURCE · 11 January 2024
Can large language models identify and correct their mistakes? ↗
What the laboratory reports
Google Research studies mistake finding separately from mistake correction.
The work treats feedback quality and repair ability as distinct capabilities rather than one self-correction score.
This decomposition exposes where a system fails: detecting an error, localizing it or producing a valid repair.
The mathematical problem
Detection, localization and repair have different loss functions. Collapsing them into a single accuracy number hides conditional failure rates, especially the probability that a wrong diagnosis leads to a persuasive but invalid repair.
GERO's proposed response
Create separate graph events for contradiction detection, faulty-node localization and dependency-safe repair.
Require a new certificate after repair; never inherit the original verdict automatically.
Track transition metrics such as invalid-to-valid, valid-to-invalid and unresolved-to-falsely-verified.
A falsifiable experiment
Inject labelled defects at different depths of a claim graph.
Score detection, localization and repair independently.
Test whether the repaired conclusion survives the same attacks as a clean reference solution.
Boundary
A model's ability to critique text does not demonstrate that its proposed correction is valid. The repaired artifact needs fresh verification.
Read the primary source
This article is an original analytical summary, not a republication. Read Can large language models identify and correct their mistakes? for the laboratory's complete claims, methods and context.
