Definition
Plain language
A simple statistical method for predicting a yes-or-no outcome from input features.
As stated in the literature
A linear model that maps features through a logistic function to a probability of a binary label; used in this corpus as a lightweight trained baseline against training-free or probe-based classifiers.
Why it matters: Its simplicity makes it a cheap, transparent baseline for judging whether fancier classifiers are actually worth their added complexity.
For example, it can predict whether an email is spam by weighing features like suspicious words and turning the total into a probability.
Heard on the show
“Beats a trained logistic regression baseline by thirteen points, beats the strongest training-free baseline by twenty-four.”Episode 023 — Why a Small Agent Confidently Overwrites Memories It Doesn't Understand