arXiv:2602.14862v2 Announce Type: replace 
Abstract: Temperature scaling is a simple method that allows to control the uncertainty of probabilistic models. It is mostly used in two contexts: improving the calibration of classifiers and tuning the stochasticity of large language models (LLMs). In both cases, temperature scaling is the most popular method for the job. Despite its popularity, a rigorous theoretical analysis of the properties of temperature scaling has remained elusive. We investigate here some of these properties. For classification, we show that increasing the temperature increases the uncertainty in the model in a very general sense (and in particular increases its entropy). However, for LLMs, we challenge the common claim that increasing temperature increases diversity. Furthermore, we introduce two new characterisations of temperature scaling. The first one is geometric: the tempered model is shown to be the information projection of the original model onto the set of models with a given entropy. The second characterisation clarifies the role of temperature scaling as a submodel of more general linear scalers such as matrix scaling and Dirichlet calibration: we show that temperature scaling is the only linear scaler that does not change the hard predictions of the model.

دراسة حديثة بعنوان 'The Well-Tempered Classifier: Some Elementary Properties of Temperature Scaling' تستكشف خصائص مقياس الحرارة، وهي طريقة تُستخدم للتحكم في عدم اليقين في النماذج الاحتمالية، وخاصة في المصنفات والنماذج اللغوية الكبيرة (LLMs). تكشف الأبحاث أن زيادة درجة الحرارة تزيد عمومًا من عدم اليقين في النموذج وزيادة الإنتروبيا، بينما تتحدى الفكرة القائلة بأنها تعزز التنوع في LLMs.

Un estudio reciente titulado 'The Well-Tempered Classifier: Some Elementary Properties of Temperature Scaling' investiga las propiedades de la escala de temperatura, un método utilizado para controlar la incertidumbre en modelos probabilísticos, especialmente en clasificadores y grandes modelos de lenguaje (LLMs). La investigación revela que aumentar la temperatura generalmente incrementa la incertidumbre y la entropía del modelo, desafiando la noción de que mejora la diversidad en los LLMs.

Une étude récente intitulée 'The Well-Tempered Classifier: Some Elementary Properties of Temperature Scaling' examine les propriétés de l'échelle de température, une méthode utilisée pour contrôler l'incertitude dans les modèles probabilistes, en particulier dans les classificateurs et les grands modèles de langage (LLMs). La recherche révèle qu'augmenter la température augmente généralement l'incertitude et l'entropie du modèle, tout en remettant en question l'idée qu'elle améliore la diversité des LLMs.

A recent study titled 'The Well-Tempered Classifier: Some Elementary Properties of Temperature Scaling' investigates the properties of temperature scaling, a method used to control uncertainty in probabilistic models, particularly in classifiers and large language models (LLMs). The research reveals that increasing temperature generally raises model uncertainty and entropy, while challenging the notion that it enhances diversity in LLMs.