A language model produces one logit for each token in its vocabulary at a generation step. A transformation such as softmax converts the logits into a probability distribution, after optional controls such as temperature, masking, repetition penalties, or watermark bias are applied.
Changing a logit changes relative token probability without rewriting the token itself. Soft watermarking exploits this property by adding a small bias to keyed preferred tokens, increasing their aggregate frequency while leaving nonpreferred choices available.
ELI5
A language model logit is a raw score for one possible next token before the scores become probabilities. Every token in the vocabulary receives a score at each generation step.
For example, raising one token's logit makes it more likely relative to alternatives after the probability conversion. Temperature, masks, penalties and watermark biases can all change these scores or how they are used.
