logit bias

Logit bias manually adjusts token probabilities before sampling to encourage or suppress specific outputs. **Mechanism**: Add (or subtract) fixed values to logits of specified tokens before softmax. Positive bias → more likely, negative bias → less likely, -100 effectively bans token. **Use cases**: Ensure specific format tokens appear, prevent problematic terms, guide structured generation, enforce vocabulary constraints. **API support**: OpenAI API accepts token ID → bias value dictionary, other providers have similar features. **Examples**: Ban curse words (negative bias), encourage JSON formatting tokens, suppress competitor names, ensure answer ends with period. **Relationship to prompting**: Complements instructions - bias provides hard constraints, prompts give soft guidance. **Tokens to bias**: Use tokenizer to find exact token IDs - be aware of multi-token words. **Trade-offs**: Can create awkward outputs if overused, may interfere with natural generation, requires knowing exact token IDs. **Best practices**: Use sparingly for critical constraints, test thoroughly, prefer prompting for soft preferences, save hard constraints for format-critical applications.