logit bias

Logit bias is a precision tool for putting your thumb on the scale of specific tokens. You supply a map from token ids to numbers, and at every generation step those numbers are added to the model's raw scores (logits) before they become probabilities. A positive bias makes a token more likely to appear; a large negative bias makes it nearly impossible. It is a surgical edit to the distribution, applied uniformly at every step the model decodes.

People reach for it to ban a profanity or a competitor's name, to force a classifier to answer only 'Yes' or 'No', or to gently encourage a domain term. The sharp edges: you must work in token ids, not words, and a word may split into several tokens, so biasing one piece does not always do what you expect. An overpowering bias can also derail fluency, since you are overriding the model's judgement everywhere. For complex constraints, grammar-based constrained decoding is usually cleaner.