Papers/2609.17560
🧪 Test?View on arXiv

Pay Only for Disagreement: Certified No-Regression Verdicts for Model Updates with Matching Label-Complexity Bounds

Author1, Author2, Author3, Author4, Author5

model auditingrisk assessmentupdate certificationlabel efficiency
2609.17560
Builder Relevance
80%
1h ago

Abstract

This paper presents a method for certifying model updates to ensure they do not regress in performance, utilizing a novel auditing protocol.

Reality Card

Core Claim

The proposed DISCERN protocol allows for the certification of benign model updates with a high degree of accuracy, achieving a miscoverage rate of only 0.0002.

Method / Result

56% of benign updates can be certified with zero labels, demonstrating significant efficiency in the auditing process.

Limitations

The method's performance may vary based on the adversarial nature of the label-routing rule used during audits.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.
← Back to all papers