Temperature scaling closes confidence gap while boosting accuracy in prompt tuning
Bridging the Confidence Gap: Temperature Scaling for Calibrating Test-Time Prompt Tuning
Machine Learning
Summary
It is hard for AI models to be both accurate and confident about their answers when they adapt to each new question. The authors found that existing fixes often reduce accuracy. They propose a simple method called CoTS that adjusts AI confidence without hurting accuracy by balancing it with the model’s original predictions. They also create a boosted version, E-CoTS, that improves accuracy and keeps confidence trustworthy across many tests.
What this means in practice
- •For machine learning engineers: Improve model confidence calibration in adaptive AI systems without losing accuracy using temperature scaling methods.
- •For computer vision developers: Enhance image classification reliability and confidence measures by integrating E-CoTS calibration for test-time prompt tuning.
Authors
Yuwei Liang, Jian Liang, Dapeng Hu, Yinuo Xu, Ran He
Abstract
Test-time prompt tuning (TPT) enables adaptation on a single test instance, achieving improved accuracy but often sacrificing calibration performance. Most existing calibration methods introduce additional regularization terms to promote dispersion across text embeddings and reduce calibration error, yet these methods often suffer from a drop in accuracy. Motivated by the well-calibrated nature of zero-shot predictions, we propose CoTS, a simple yet effective post-hoc calibration method that preserves accuracy. Specifically, CoTS applies temperature scaling to minimize the confidence gap between adapted and zero-shot predictions. To fully exploit the potential of multiple augmentations during adaptation, we introduce a weak-strong ensemble strategy that further boosts accuracy. We then apply CoTS to this ensemble, termed E-CoTS, to maintain its well-calibrated property. Extensive experiments on diverse datasets and backbones show that our approaches effectively mitigate miscalibration without compromising primary accuracy. For instance, E-CoTS reduces the average expected calibration error of TPT from 11.90% to 5.38% on ImageNet variants, while even increasing accuracy from 60.74% to 62.95%. Moreover, when integrated with existing calibration methods, E-CoTS usually enhances both accuracy and calibration simultaneously.