Prediction-preserving repair separates confidence correction from top-1 decisions

CORD is a post-fit adapter that repairs calibrated probability vectors so the original top-1 prediction remains the argmax, achieving zero Top-1 Prediction Change Rate (TPCR) by construction on CIFAR-10/100 and ImageNet-1K while lowering mean ECE, NLL, and Brier score. The approach moves the decision-preservation constraint out of calibrator fitting and into output-level repair, requires no additional supervised model, and has no user- or validation-tuned hyperparameters. The source is an arXiv abstract with method details still to be verified.

Hugging Face Daily Papers — Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration

September 5, 20261 source73% confidence

72Strategic

Strategic score

Evidence strength47

Technical novelty82

Commercial potential56

This source announces CORD (Calibrator-Output Repair for Top-1 Decision Preservation), a post-fit adapter that repairs the full calibrated probability vector so that the original top-1 class remains the argmax. The claimed guarantee is exact decision preservation: CORD reports zero Top-1 Prediction Change Rate (TPCR) by construction on CIFAR-10/100 and ImageNet-1K, while also reducing mean ECE, NLL, and Brier score relative to the direct calibrated output.¹

Previously, prediction flips were at best absorbed into aggregate accuracy because accuracy couples prediction changes with correctness changes. CORD relocates the preservation requirement from the calibration objective to an output-repair step; it does not alter the fitted calibrator or its direct output and introduces no user-tuned or validation-tuned hyperparameters.¹

Context

A multiclass calibrator optimized for confidence error can still change which class is top-1, and accuracy masks the frequency of those flips. CORD is positioned as the first method to impose exact top-1 preservation through post-fit vector repair rather than through constrained fitting or an additional supervised map. This design makes reliable confidence and stable decisions separable objectives instead of competing ones.¹

Assessment

The evidence is consistent but compressed: across all three vision benchmarks the source reports lower mean ECE, NLL, and Brier alongside zero TPCR, and it claims paired gains persist under distribution shift and across calibration-set sizes. What is missing from this abstract are error bars, effect magnitudes, comparison methods, and a full derivation of the repair guarantee, so the strength of current evidence is moderate rather than conclusive.¹

Caveats

AI hypothesisI would change my assessment if full-text inspection shows CORD can fail to preserve the original top-1 on some inputs, or if the reported metric gains arise from an artifact of vector renormalization rather than from the repair mechanism. The claim would also be weakened if the method effectively needs per-input information beyond the original and calibrated outputs, which the abstract does not spell out.¹

Source text

Hugging Face Daily Papers — Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration · September 2, 2026

Computer Science > Machine Learning

Title:Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration

Abstract:Post-hoc calibration corrects reported confidence, yet a multiclass calibrator can also change the associated top-1 prediction. Accuracy captures only the net effect of these changes on correctness, not how often predictions change; the Top-1 Prediction Change Rate (TPCR) instead measures this frequency. We propose Calibrator-Output Repair for Top-1 Decision Preservation (CORD), the first post-fit adapter to impose exact prediction preservation by repairing the full calibrated probability vector. From the original and calibrated outputs alone, CORD determines the mass assigned to the original top-1. The calibrated conditional distribution allocates the remaining mass over the other classes, yielding a repaired vector whose own argmax recovers the original prediction. On the calibration split, CORD coordinates the repaired masses to retain the calibrated outputs' mean mass on original predictions whenever attainable. The adapter alters neither the fitted calibrator nor its direct output, fits no additional supervised map, and requires no user- or validation-tuned hyperparameter. Across CIFAR-10/100 and ImageNet-1K, CORD attains zero TPCR by construction and lowers mean ECE, NLL, and Brier relative to the corresponding direct outputs in every dataset; paired gains persist under distribution shift and across calibration-set sizes. CORD thus removes the preservation constraint from calibrator fitting and assigns exact recovery of the original decision to subsequent output repair. Our code is available at this https URL.

Access Paper:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

A point needing decoding is how zero TPCR can hold for arbitrary calibrated outputs. Merely taking the mass assigned to the original top-1 and renormalizing the remaining conditional distribution would leave the original class as argmax only when it already receives the largest mass; otherwise some extra mass must be granted to that class. The abstract's wording suppresses this lower-bound/threshold operation, and the proof of zero TPCR depends on it.

The phrase 'retain the calibrated outputs' mean mass on original predictions whenever attainable' is an aggregate condition, not a per-input guarantee. It allows individual cases to receive more or less mass than the direct calibrated output as long as the calibration-split average is preserved, and the word 'attainable' signals a feasibility condition that the abstract leaves undefined.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration

September 2, 2026endistinct angle

Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration

↑ Back to top