Adjust a model's scores for the next word and watch cross-entropy loss and perplexity respond — the exact quantity language models are trained to minimize.