A logit is not a confidence score. Learn how logits become probabilities, why a high probability is not always confidence, and what calibration fixes.
Updated August 2026
A machine learning model can report that it is 99% sure and still be wrong far more than 1% of the time. That gap comes from treating three different things, logits, probabilities, and confidence, as one.
Quick answer: A logit is the raw score a model outputs before it becomes a probability. Softmax or sigmoid turns logits into probabilities between 0 and 1. A probability looks like a confidence score, but it is only trustworthy if the model is calibrated. Modern neural networks are often overconfident, so a reported 99% can mean much less than 99% accuracy.
These three sit in a chain, and each step changes what the number means.
This content was generated with the assistance of AI. Our AI prompt chain workflow is carefully grounded and preferences .gov and .edu citations when available. All content is reviewed by a Telnyx employee to ensure accuracy, relevance, and a high standard of quality.
The mistake is to collapse the three into one, reading a large logit or a high probability as proven certainty. The size of a logit tells you which class the model prefers, not how often that preference is right.
Softmax turns a vector of logits into a probability distribution. It exponentiates each logit and divides by the sum of all of them, so larger logits get larger probabilities and the values add to 1. PyTorch's softmax does exactly this.
For a single binary output, sigmoid plays the same role, mapping one logit to one probability. Either way, the probability is a rescaling of the logits, not a new measurement of how correct the model is.
Yes, they match. Softmax is monotonic: it never changes the order of the values it transforms, only their scale. The largest logit always produces the largest probability, so the predicted class, the argmax, is identical whether you take it from the logits or from the probabilities. This is why inference code often skips softmax and takes the argmax of the raw logits directly.
The order is preserved; only the interpretation changes. Logits tell you the ranking, and softmax adds a scale that looks like probability.
A high probability is not a reliable sign of confidence. It looks like confidence, but whether it earns that reading depends on calibration. Guo and colleagues showed that modern neural networks are often overconfident: a model that outputs 99% can be right well below 99% of the time, and this miscalibration is worse in modern deep networks than in the shallower ones of a decade earlier.
Picture a classifier that outputs 0.95 across a batch of predictions. If it is well calibrated, about 95 of every 100 such predictions are correct. An overconfident model might get only 75 of them right while still reporting 0.95: the same number on the screen, a very different reality behind it.
The cause is that softmax will happily produce a sharp, near-certain distribution even when the model has no basis for it. The probability reflects the relative size of the logits, not the model's true hit rate. Treating it as a literal confidence score is where risk decisions go wrong.
The same choices that make modern networks accurate also make them overconfident. Large capacity lets a model drive its training loss toward zero, which pushes its softmax outputs toward 0 or 1 regardless of real certainty. Long training schedules and some normalization techniques add to the effect.
The result is a model that is usually right and almost always sure, even when it should not be. Accuracy and calibration are separate properties. Improving one does not automatically improve the other, which is why calibration has to be checked on its own.
Calibration is the alignment between a model's stated confidence and its actual accuracy. A well-calibrated model that says 80% is right about 80% of the time. Model calibration is how you check and correct that alignment.
Temperature scaling, from the same work, is a simple fix. It divides the logits by a single learned number before softmax, softening an overconfident distribution. Because it only rescales the logits, it leaves the argmax, and therefore the predictions, untouched. It changes the confidence, not the answer.
You measure it by comparing predicted confidence with observed accuracy. Group predictions by their confidence, then check how often each group is actually correct. A reliability diagram plots the two: a perfectly calibrated model follows the diagonal, while an overconfident one sits below it, with accuracy trailing confidence.
Expected calibration error, or ECE, condenses that gap into a single number, the average distance between confidence and accuracy across the groups. A lower ECE means the probabilities can be trusted more directly. The point is that calibration is measurable, so "is this confidence real" is a question you answer with held-out data rather than assume.
Treat a raw probability as a ranking signal first and a trustworthy rate second. It is fine for choosing the top class or ordering predictions from most to least likely. It is not safe as a literal probability, for thresholds or risk decisions, until you have checked calibration on held-out data.
When the decision matters, measure calibration and recalibrate if needed. Pair the confidence with the metric you actually care about, such as the F1 score, rather than the model's self-reported certainty. This matters most at inference, where the score drives a real action.
No, a logit is a raw, unbounded score a model produces before any probability function. It becomes a probability through softmax or sigmoid, and only becomes a trustworthy confidence score if the model is calibrated. A large logit signals a preferred class, not a verified level of certainty.
Softmax converts a vector of logits into probabilities by exponentiating each one and dividing by their sum, so the values fall between 0 and 1 and add to 1. For a single binary output, sigmoid maps one logit to one probability. The larger the logit, the larger the resulting probability.
A high softmax probability does not reliably mean confidence. Softmax can produce a near-certain probability even when the model is wrong, and modern networks are often overconfident. A high probability is trustworthy as a confidence score only when the model is calibrated, meaning its stated confidence matches its actual accuracy.
Calibration is the match between a model's stated confidence and how often it is correct. A calibrated model that predicts 80% is right about 80% of the time. Temperature scaling, which rescales the logits by a single learned value, is a common way to correct an overconfident model.
Yes, they are equal. Softmax preserves order, so the largest logit always maps to the largest probability. The predicted class is the same whether you take the argmax of the logits or of the probabilities, which is why inference code often skips softmax and uses the logits directly.
