Self-correction is two problems, not one

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.

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.