Logits

A language model, meaning a program that continues text, has to make a decision at every step: which word comes next? To do this, it scores every possibility it knows about with a single number each. These raw scoring numbers are called logits. They are not yet percentages: a logit can be 12.4, but it can also be minus 3.8. What matters is only the comparison between them. Whoever has the largest logit is the model’s favorite.

What logits reveal about the model’s uncertainty

Logits are the last point at which you can still watch the model 'think' in an unfiltered way. After that, only a single word gets selected, and all other information disappears. Anyone reading the logits, by contrast, sees not just the favorite but also the runner-up positions down to number ten.

That is exactly what makes them useful for measuring uncertainty. If the best logit sits far ahead of all the others, the model is very confident. If five values are nearly tied, it is essentially a coin flip. For a legal question or a medical statement, this distinction matters more than the answer itself. Some systems therefore flag such answers as unreliable or ask a follow-up question.

A common misconception: a high logit does not mean the statement is true. It only means that this word fits well with the patterns of the training texts. A model can therefore be wrong with great confidence. Certainty and correctness are two different things.

From raw number vector to probability

At the end of the network sits a long list of numbers, one for every known piece of text. For large models, that is roughly 100,000 to 200,000 entries. This list is the logit vector. It gets converted using a fixed formula called softmax. Softmax amplifies large values, pushes small ones down, and ensures that in the end all the numbers together add up to exactly 100 percent.

Between logits and softmax, a dial often comes into play: temperature. You divide all the logits by a number. Dividing by a small value makes the gaps larger and the model answers very predictably. Dividing by a large value pulls the possibilities closer together and the answers become more varied, but also riskier. The same logits can thus lead to quite different texts.

An analogy helps here: logits are like scores from judges at a gymnastics competition. The raw points don’t yet say who has what percentage chance of winning. Only the conversion turns them into comparable shares. And whoever stretches or compresses the scale changes the result without touching a single score.

Logits in interfaces, tools, and headlines

Anyone working with a model through a programming interface can often have the so-called logprobs output. These are the probabilities calculated from the logits, expressed in a special notation. Developers use them to automatically check how confident an answer was. Tools that try to detect machine-generated text also work with such values.

In research, logits come up when shrinking models. In so-called distillation, a small model learns not just the correct answer from the large one, but its entire scoring list. This extra information makes the small model considerably better than if it only saw the final results. Such compact models then run on phones instead of in data centers.

In the news, the term usually appears indirectly. When it’s reported that a provider has disabled logprobs, it’s about control: these values reveal a lot about a model’s internal structure. And when a provider advertises particularly creative or particularly precise answers, often what’s really behind it is simply a different temperature applied to the same logits.