Definition
Plain language
The raw numbers a model produces before they get turned into probabilities over words.
As stated in the literature
The unnormalized scores output by a model's final layer, converted to a probability distribution via softmax.
Also called: logit
Why it matters: Operating on logits rather than probabilities — for sampling, calibration, or distillation — is standard practice and avoids precision issues.
For example, a model might output logits of (2.1, 0.4, -1.0) over three tokens, which the softmax then turns into probabilities of roughly (0.78, 0.14, 0.08).
Heard on the show
“For the frontier models, where they can't read logits and have to sample outcomes, candidates get screened on one batch and confirmed on another.”Episode 243 — How a Hundred Meaningless Word Choices Add Up to Flip a Model's Answer