Logits

The raw, unnormalized output scores produced by a neural network's final layer before conversion to probabilities. In language models, logits represent confidence for each possible next token. Sampling strategies like temperature and top-k operate on these values.