Logit Bias

An inference-time intervention technique applied during the decoding stage of large language models. It adds a fixed bias value to the raw scores (logits) of specified tokens before the softmax operation, adjusting their generation probabilities to influence output distribution without modifying model weights.