LiST: A New Method Delivers Accuracy, Robustness and Calibration in Neural Networks at Once

Researchers published LiST (Lipschitz Scaling Training) on arXiv in July 2026. The new approach to training neural networks delivers accuracy, robustness to adversarial attacks and calibration at the same time, without manual parameter tuning.

Three properties that are hard to combine

A reliable neural network must satisfy three conditions at once. Accuracy: the model correctly classifies ordinary data. Robustness: it does not break when input data is deliberately and imperceptibly distorted (adversarial attacks). Calibration: the model's stated confidence in its predictions matches its actual accuracy. If the model says "80% probability", it is right in roughly 80% of cases.

The three properties are usually studied separately, and improving one often degrades another. The existing class of Lipschitz-constrained models handles robustness well: the Lipschitz constant L limits how much the network's output changes when the input changes slightly. However, the right value of L has traditionally been chosen by hand, and its effect on calibration has remained almost unexplored.

How does LiST work?

The authors found a theoretical and empirical link between the Lipschitz constraint and Temperature Scaling, a popular post-processing method for calibrating neural networks. The main result: for any training scheme there is a specific value L* at which the network is automatically calibrated, with no additional steps. Moreover, calibration can serve as a principled criterion for choosing an operating point on the accuracy–robustness Pareto front.

LiST iteratively adjusts the global Lipschitz constant during training until the model reaches L*. The margin parameter in the loss function makes it possible to build the full calibrated Pareto front: the user gets a set of models with different balances of accuracy and robustness, each of which stays calibrated by default. Once training converges, LiST allows the calibration data to be added back into training, which improves sample efficiency without losing calibration.

Key facts from the study:

• Testing was done on CIFAR-10, CIFAR-100 and Tiny-ImageNet
• Calibration is achieved "out of the box", without post-processing
• The method supports building the full accuracy–robustness Pareto front
• Adding the calibration data back into training improves sample efficiency
• The code is available on GitHub

What this means

LiST offers a systematic way to tackle three key neural network problems at once, without compromises and without manual hyperparameter search. For ML engineers, this shortens the model tuning cycle before deployment. For researchers, it provides a principled tool for managing the balance between accuracy and robustness, with guaranteed calibration across the whole front.

Frequently asked questions

What is neural network calibration, and why does it matter?

A calibrated model is one whose stated confidence matches the actual frequency of correct answers. If such a model predicts "80% probability of class A", it is right about 80% of the time. This is critical in medicine, finance and security, where incorrect probabilities lead to costly mistakes.

How is LiST related to Temperature Scaling?

Temperature Scaling divides the model's logits by a constant T, which changes the "sharpness" of the probability distribution. The LiST authors showed that the Lipschitz constraint affects accuracy and robustness in a way similar to how T affects calibration, and this is what lets them merge both mechanisms into a single training scheme.