Calibration vs Decision Making: Revisiting the Reliability Paradox in Unlearned Language Models

2026-05-20

AI Summary

This study reveals that language models, despite exhibiting well-calibrated outputs after machine unlearning (ECE ≈ 0.04), still rely on spurious correlations in their decision-making, undermining their reliability. Through multiple-choice question answering on the TOFU benchmark, the authors evaluate calibration using Expected Calibration Error (ECE), Maximum Calibration Error (MCE), and Brier score, while employing Integrated Gradients and local mutual information for attribution analysis. This work extends the “reliability paradox” to the machine unlearning setting for the first time, demonstrating that low calibration error can mask dependence on shortcut features. The findings highlight a disconnect between good calibration and trustworthy reasoning, challenging the common assumption that calibration serves as a reliable proxy for model trustworthiness.

Abstract

Machine unlearning aims to remove the influence of specific training data from a model while preserving reliable behavior on the remaining data, making reliable prediction and uncertainty estimation essential for evaluation. Calibration is commonly used as a proxy for reliability in language models, but low calibration error does not necessarily imply reliable decision rules, as models may rely on spurious correlations while remaining well calibrated. We investigate this gap in generative language models using the multiple-choice question-answering evaluation protocol on the TOFU benchmark, measuring probabilistic reliability with calibration metrics (ECE, MCE, Brier) and decision-rule reliability via attribution-based shortcut detection with Integrated Gradients and Local Mutual Information. We find that fine-tuned models achieve low calibration error (ECE ~ 0.04) compared to pretrained models (ECE > 0.5), and models after unlearning retain similarly low calibration despite reduced accuracy on the forget split, while attribution analysis shows increased reliance on correlation-based tokens. These results demonstrate that good calibration can coexist with shortcut-based decision rules after unlearning, extending the reliability paradox to the machine unlearning setting.

Problem

• machine unlearning
• calibration
• reliability paradox
• shortcut learning
• language models

Innovation

• machine unlearning
• calibration
• reliability paradox
• attribution analysis
• shortcut learning