logit-bias

This paper proposes a black-box method for adapting language models by learning a single context-independent logit-bias vector added at every decoding step, which improves performance on math and reasoning tasks without modifying model weights or requiring gradients.