Soft Labels
Soft labels are training targets that assign an example not just a single correct answer, but different probabilities across several possible answers. Instead of "this is a husky," the target becomes "80 percent husky, 15 percent wolf, 5 percent German shepherd."
For a computer program to learn to classify images or texts, it needs training examples paired with the respective correct answer. This attached answer is called a label in technical terms. Normally, a label is unambiguous: this image shows a husky, period. With soft labels, however, the answer is split up. The example is considered 80 percent husky, 15 percent wolf, and 5 percent German shepherd. The hard either-or specification is thus replaced by a distribution that can express uncertainty.
What unambiguous answers conceal
The world is rarely unambiguous. A blurry photo might really show a husky, but look confusingly similar to a wolf. An unambiguous label nevertheless claims: one hundred percent husky, zero percent wolf. From this, the program learns a certainty it cannot actually have.
Such exaggerated certainty is a well-known problem. Experts speak of overconfidence: the model states 99 percent certainty for its answer and is still wrong. Anyone relying on this number, say in medicine or in automatic document review, is being misled. Soft labels dampen this effect, because the model learns from the outset that some cases are ambiguous.
A second advantage is the additional information. The statement “15 percent wolf” reveals which classes resemble each other. A hard label only says what is wrong, but not how wrong. That’s why models trained on soft labels often learn the same thing with fewer examples.
Where the percentages come from
One source is humans. If you have twenty people classify the same image and sixteen vote husky while four vote wolf, this directly produces a soft label of 80 to 20. The disagreement among raters is thus not discarded but turned into learning material. In classic datasets, the majority vote is taken instead and the rest is discarded.
The more common source is another model. A large, expensive model is made to evaluate all training examples, and its probabilities are recorded as new labels. A small model is then trained on these values. This procedure is called knowledge distillation, because the knowledge of the large model is transferred into a smaller one. The large model is called the teacher in this process, the small one the student.
There is also a simple variant called label smoothing. Here, you take a hard label and distribute a small fraction, say one tenth, evenly across all other classes. 100 percent husky then becomes 90 percent husky and a little bit of everything else. This costs almost nothing and measurably reduces overconfidence. However, it should not be confused with genuine soft labels, since it contains no additional information about similarities.
Soft labels in today’s AI products
Soft labels are most visible in the smaller variants of large language models. Almost every provider sells a faster, cheaper version alongside its flagship model for mass deployment. These small models are usually created through distillation, that is, by training on the soft outputs of the larger sibling model. This is precisely why some assistants today run on a phone.
In image processing, soft labels appear with uncertain images. A system for detecting skin lesions doesn’t output “malignant,” but a probability. Doctors can use this to decide whether further examination is needed. A model that only spits out hard verdicts would be useless for this purpose.
In the news, you mostly encounter the term as a point of contention. When a company trains a model on the outputs of someone else’s model, this often violates the provider’s terms of use. Such accusations surrounding distillation repeatedly made headlines in 2025. The harmless technical core behind them is precisely these soft target values.