LiST: a new method simultaneously delivers accuracy, robustness, and calibration in neural networks

Researchers in July 2026 published on arXiv the LiST (Lipschitz Scaling Training) method — a new approach to training neural networks that simultaneously ensures accuracy, robustness to adversarial attacks, and calibration without manual hyperparameter tuning.

Three Properties That Are Hard to Reconcile

A reliable neural network must satisfy three conditions simultaneously. Accuracy — the model correctly classifies ordinary data. Robustness — it does not fail when input data is intentionally and imperceptibly distorted (adversarial attacks). Calibration — the model's claimed confidence in its predictions matches real accuracy: if the model says "80% probability," then in ~80% of cases it is actually correct.

All three properties are traditionally studied separately, and improving one often worsens another. An existing class of models with Lipschitz constraints handles robustness well: the Lipschitz constant L constrains how much the network output changes with small input changes. However, the required L value was traditionally chosen manually, and its effect on calibration remained practically unexplored.

How LiST Works

The authors discovered a theoretical and empirical connection between Lipschitz constraints and Temperature Scaling — a popular post-processing method for neural network calibration. The main result: for any training scheme, there exists a specific value L* at which the neural network automatically becomes calibrated — without additional steps. Moreover, calibration can be used as a principled criterion for selecting a working point on the Pareto front of "accuracy–robustness."

The LiST method iteratively adjusts the global Lipschitz constant during training until the model reaches value L*. The margin parameter in the loss function allows constructing a complete calibrated Pareto front: the user obtains a set of models with different accuracy-robustness trade-offs, each of which remains calibrated by default. Upon convergence, LiST allows re-including calibration data in training — this increases sample efficiency without losing calibration.

Key facts from the research:

• Testing was conducted on CIFAR-10, CIFAR-100, and Tiny-ImageNet
• Calibration is achieved "out of the box," without post-processing
• The method supports building a complete Pareto front of accuracy–robustness
• Re-including calibration data in training improves sample efficiency
• Code is available on GitHub

What This Means

LiST offers a systematic way to address three key neural network problems simultaneously — without compromises and without manual hyperparameter search. For ML engineers, this reduces 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 entire front.

Frequently Asked Questions

What is neural network calibration and why is it needed?

A calibrated model is one where the claimed confidence matches the actual frequency of correct answers. If such a model predicts "80% probability of class A," then in approximately 80% of cases it is correct. This is critical in medicine, finance, and security, where incorrect probabilities lead to costly errors.

How is LiST related to Temperature Scaling?

Temperature Scaling divides the model's logits by a constant T, which changes the "steepness" of the probability distribution. The LiST authors showed that Lipschitz constraints affect accuracy and robustness similarly to how T affects calibration — this enables combining both mechanisms into a single training scheme.