Logit bias is a technique that allows users to influence the output of language models by adjusting the likelihood of specific tokens being selected. This method modifies the logits, which are scores assigned to potential next tokens, before they are converted into probabilities. By applying a positive bias, a token becomes more favored, while a negative bias makes it less likely to be chosen, with a strong negative bias capable of preventing its selection entirely. This offers a way to control token generation without altering the model's underlying parameters. AI
IMPACT Logit bias provides a method for fine-tuning LLM output without retraining, enabling more precise control over generated text.
RANK_REASON The item describes a technique for controlling LLM output, which is a tool or method rather than a core AI release or research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
