Tag
论文提出“watermarkability”指标来量化语言模型与基于logit的水印规则的兼容性，并通过两阶段方法（目标集条件LoRA加闭环控制器）在保持高水印可靠性的同时显著降低困惑度增加。
This paper proposes a black-box method for adapting language models by learning a single context-independent logit-bias vector added at every decoding step, which improves performance on math and reasoning tasks without modifying model weights or requiring gradients.