Adjustment to token logits at inference time to steer generation or block outputs.