Learning When to Be Uncertain: A Post-Hoc Meta-Model for Guided Uncertainty Learning

ICLR 2026 Conference Submission11279 Authors

Keywords: Uncertainty Quantification, Adversarial Robustness, OOD Detection

TL;DR: GUIDE is a lightweight plug-in evidential meta-model that teaches frozen deep networks when to be uncertain, improving OOD and adversarial detection without retraining or architectural changes.

Abstract: Reliable uncertainty quantification remains a major bottleneck in deploying deep learning models under distribution shift. Existing methods that retrofit pretrained models either inherit misplaced confidence or merely reshape predictions, without teaching the model when to be uncertain. We introduce GUIDE, a lightweight evidential learning meta-model approach that attaches to a frozen deep learning model and explicitly learns how and when to be uncertain. GUIDE identifies salient internal features via a calibration stage, and then employs these features to produce a noise-driven curriculum that teaches the model how and when to express uncertainty. GUIDE requires no retraining, no architectural modifications, and no manual intermediate-layer selection to the base deep learning model, making it broadly applicable while reducing user effort. The resulting model avoids distilling overconfidence from the base model, improves out-of-distribution detection ($\approx$ 77\%) and adversarial attack detection ($\approx$ 80\%), as well as preserving in-distribution performance. Across diverse benchmarks, GUIDE consistently outperforms state-of-the-art approaches, demonstrating that actively guiding uncertainty is key to closing the gap between confidence and reliability.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 11279