Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration
Summary
This paper proposes CORD, a post-fit adapter for post-hoc calibration that repairs probability vectors to preserve original top-1 predictions while maintaining calibration quality, achieving zero TPCR and improved metrics on datasets like CIFAR and ImageNet.
View Cached Full Text
Cached at: 09/04/26, 07:56 AM
Paper page - Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration
Source: https://huggingface.co/papers/2609.01072
Abstract
CORD is a post-fit adapter that repairs calibrated probability vectors to exactly preserve original top-1 predictions while maintaining calibration quality.
Post-hoc calibrationcorrects reported confidence, yet amulticlass calibratorcan also change the associated top-1 prediction. Accuracy captures only the net effect of these changes on correctness, not how often predictions change; theTop-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 fullcalibrated probability vector. From the original and calibrated outputs alone,CORDdetermines 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 ownargmaxrecovers the original prediction. On the calibration split,CORDcoordinates 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,CORDattains zero TPCR by construction and lowers meanECE,NLL, and Brier relative to the corresponding direct outputs in every dataset; paired gains persist underdistribution shiftand across calibration-set sizes.CORDthus removes the preservation constraint from calibrator fitting and assigns exact recovery of the original decision to subsequent output repair. Our code is available at https://github.com/labhai/CORD.
View arXiv pageView PDFGitHubAdd to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.01072 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.01072 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.01072 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Calibration-Preserving Pruning: Compression as a Reliability Contract
The paper introduces Calibration-Preserving Pruning (CPP), a method to enhance pruning for maintaining conformal prediction efficiency, showing reduced prediction set sizes and improved accuracy on benchmarks like DBpedia-14 with models such as Qwen2.5-1.5B.
FALCON-Discover: Discovering Concentrated False-Confidence Regions for Calibration
This paper introduces FALCON-Discover, a post-hoc framework for discovering regions where model predictions are confidently wrong, shifting calibration focus from aggregate metrics to sample-level dangerous failures.
Retrieval-Corrected Conformal Prediction for Time Series
Introduces Retrieval-Corrected Conformal Prediction (RCCP), a retrieval-augmented calibration method for time series prediction intervals that selects similar past residuals and applies a scalar conformal correction to achieve target coverage with low overhead.
Expectation Consistency Loss: Rethink Confidence Calibration under Covariate Shift
This paper introduces the Expectation Consistency Loss (ECL), a theoretically grounded loss function for calibrating classifier confidence under covariate shift, derived from a necessary and sufficient condition called the Expectation Consistency Condition.
The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation
This paper identifies that on-policy distillation (OPD) in language models leads to severe overconfidence due to information mismatch between training and deployment, and proposes CaOPD, a calibration-aware framework that improves both performance and confidence reliability.