Logits

The last step hands out one score per candidate

The model keeps reshaping values as a sentence passes through it, then runs one more calculation at the very end to produce a row of numbers. That row is exactly as long as the number of pieces the model knows — tens of thousands of pieces means tens of thousands of scores.

Each entry stands for "how likely this piece is to come next." A larger value means the model favors that piece; a small or negative one means it doesn't fit the current spot.

These numbers aren't written for a person to read. They're a direct byproduct of computation happening inside the model, with no fixed scale or reference point of their own.

The scores aren't probabilities yet

Logits differ from probabilities in three ways: they don't sum to one, they can be negative, and they have no upper or lower bound. So one score alone can't tell you "how confident is the model in this answer."

What matters is the gap between candidates. A wide gap between first and second place means the answer has essentially settled; scores bunched close together means several outcomes would all be unsurprising. Add the same amount to every score and this gap stays exactly the same, so the result doesn't change either.

Turning scores into probabilities is a separate step

Turning the scoreboard into probabilities means dividing the whole set at once. One candidate's score alone can't tell you a percentage — the scores every other candidate received matter too. If the rest score low, the same score becomes a high probability; if everyone scores similarly high, the probability splits thin.

This conversion never touches the ranking. Whoever led on the scoreboard still leads after becoming a probability. Probability isn't built to set the order — it's built to give people a scale they can read and compare, and it's also the form needed to randomly pick an answer and add some variety.

Most dials reach into this scoreboard

Settings that shape an answer's character usually step in before scores turn into probabilities. Raising temperature narrows the gap between candidates' scores, so once turned into probabilities, lower-ranked candidates become plausible picks too. Lowering it widens the gap, making the top candidate nearly a lock.

Blocking a specific word from appearing works here too — push that piece's score down far enough and it's effectively zero once it becomes a probability. Discouraging repeated words works the same way, by shaving a little off the score of pieces that already appeared.

That makes it clear where to look when an answer goes wrong: whether the model's own scores are the problem, or whether an adjustment tacked on afterward is.