Logit Bias
A per-token nudge added to the model’s logits to make specific tokens more or less likely.
Logit bias lets you add a fixed offset to the logits of specific tokens before sampling, pushing the model toward or away from them — a coarse steering knob that operates below the level of the prompt.
Worked example: to force a yes/no answer you can heavily boost the tokens for “Yes” and “No” and ban the rest; to stop a model ever emitting a particular token you set its bias to a large negative number, effectively banning it. Gotcha: bias is applied per-token, not per-word, so a word may span several tokens (biasing only the first does not ban the word), and a large positive bias can force unnatural, broken output — it is a blunt instrument, best for small, well-defined vocabularies.