Soft language model watermarking adds a controlled bias to the logits of green-listed tokens before sampling. A strongly preferred factual continuation can still win even when it is not green, while uncertain choices are gently shifted toward tokens that contribute to the hidden pattern.
This approach preserves quality better than a hard ban but creates a weaker signal. Reliable detection usually needs enough tokens and a statistical test that compares the observed green-token count with the amount expected under unwatermarked sampling.
