Logits

What it is

Logits are the raw numerical outputs from a neural network before they're converted to probabilities. For language models, logits represent the model's score for each possible next token. Higher logits mean the model considers that token more likely. Logits are converted to probabilities using softmax.

Why it matters

Logits are mostly technical, but understanding them helps grasp how models make decisions. When AI picks "surprising" words, it's because those tokens had higher logits than expected. Some advanced users analyse logits to understand model confidence or detect unusual behaviour.