Inference
logit
A logit is an unnormalized score for a vocabulary item before softmax turns it into a probability.
Expanded definition
logit shows up constantly when teams ship LLM features. Practically, it influences how you design prompts, evaluate quality, and reason about failure modes. Teams should document how logit manifests in their stack—data handling, evaluation, and runtime guardrails—and revisit assumptions as models update.
Related terms
Explore adjacent ideas in the knowledge graph.