LiST training achieves robust, calibrated neural networks with adjustable Pareto front

Neural networks usually trade off accuracy, robustness, and calibration. A robust model often becomes overconfident or underconfident, requiring post-hoc calibration techniques like Temperature Scaling. LiST (Lipschitz Scaling Training) from researchers at IRIT, Université de Toulouse, and SNCF solves this by theoretically and empirically linking the enforced Lipschitz constraint to temperature scaling. They show there exists a non-trivial Lipschitz value L* that yields a naturally calibrated network. LiST iteratively adjusts the global Lipschitz constant during training to reach this L* operating point, and introduces a margin parameter in the loss to construct a fully calibrated Pareto front. This lets users dial between accuracy and robustness while keeping the network calibrated at every step.

On CIFAR-10, CIFAR-100, and Tiny-ImageNet, LiST achieves competitive accuracy and robustness compared to both Lipschitz-constrained and standard unconstrained models, while staying calibrated without any post-processing. The method also allows calibration data to be reintegrated into training, improving sample efficiency without harming calibration. Code is available on GitHub. This work takes a major step toward reliable AI systems that don't force practitioners to choose between correct, safe, and well-calibrated outputs.