---
title: Logit
description: 'An unnormalized model score supplied to a normalization function such as softmax; in statistics, logit also means the logarithm of odds.'
canonical_url: 'https://darkfactory.dev/glossary/logit'
markdown_url: 'https://darkfactory.dev/glossary/logit.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Logit


## Definition

In modern machine learning, a logit is an unnormalized score produced for a candidate class or next token before a function such as softmax converts a vector of scores into a probability distribution. A logit can be any real number. Its absolute value is less informative than its differences from the competing logits.

Softmax exponentiates the scores and normalizes them so the outputs fall between zero and one and sum to one. Adding the same constant to every logit leaves the resulting probabilities unchanged. Raising one logit relative to the others raises that candidate's probability.

## Origin of the term

Joseph Berkson introduced "logit" in his 1944 paper *Application of the Logistic Function to Bio-Assay*. He used it for the natural logarithm of the odds, by analogy with the earlier term "probit." That statistical meaning remains exact for binary logistic models.

Neural-network practice broadened the plural "logits" to cover the raw score vector before normalization, especially in multiclass classification and language modeling. Those scores are not individually literal log-odds against every alternative. The older statistical definition and the current engineering usage are related but should not be treated as identical.

## Operational significance

Decoding systems can modify logits before selecting a token. Temperature scaling, repetition penalties, token bans, grammar constraints, and tool masks all change which outputs remain likely or possible. The Manus production account describes masking logits to restrict available actions without rewriting tool definitions in the prompt.

Logits are not calibrated confidence values. A model can assign the largest logit to an incorrect token, and a softmax probability can be sharp even when the model is wrong.

## Distinguish it from nearby terms

- Probability is normalized across the alternatives under consideration.
- Log probability is the logarithm of a normalized probability. A raw machine-learning logit is usually computed before that normalization.
- Confidence is an interpretation of probability that requires calibration against observed outcomes. Softmax alone does not provide that calibration.

## Check your understanding

Two systems add 100 to every logit before applying softmax and produce the same probabilities. Explain why the unchanged differences, rather than the absolute scores, determine the result.

## Related terms

- [Probability distribution](https://darkfactory.dev/glossary/probability-distribution)
- [Sampling](https://darkfactory.dev/glossary/sampling)

## Evidence and further reading

- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)
- [Application of the Logistic Function to Bio-Assay](https://doi.org/10.1080/01621459.1944.10500699)
- [PyTorch: Softmax](https://docs.pytorch.org/docs/stable/generated/torch.nn.Softmax.html)
- [Context Engineering for AI Agents: Lessons from Building Manus](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus)