研究人员提出了一种名为PTP(基于前一个令牌预测的大语言模型反转)的新方法,旨在实现近乎精确的提示词重构 1。该技术的核心机制是在无需访问目标大语言模型权重或外部辅助数据的情况下,通过在目标模型合成的数据上从头训练一个显式的逆语言模型来工作 1。PTP采用“前一个令牌预测”作为训练目标,这与正向语言模型常用的“下一个令牌预测”形成类比 1。
实验结果显示,PTP在黑盒设置下具备跨数据集的泛化能力和可转移性 1。在多项评估指标中,该方法的表现优于现有的相关工作 1。此外,PTP还支持通过采样生成多种能够诱导相似响应的提示词 1。根据相关论文信息,其提交日期标注为2026年7月31日 1。
A new study presents a method called Previous-Token Prediction based LLM Inversion (PTP), designed to reconstruct prompts with near-exact precision 1. The core mechanism involves training an explicit inverse language model from scratch using data synthesized by the target large language model, rather than relying on pre-existing models or weights 1. By utilizing previous-token prediction as its training objective—mirroring the standard next-token prediction process used in forward language modeling—the approach achieves high fidelity in reversing the generation process 1.
The research emphasizes that PTP operates strictly within a black-box setting, requiring no access to model weights, logits, or external fine-tuning datasets 1. Experimental results indicate that the method demonstrates strong generalization and transferability across different datasets while outperforming previous works on token-based evaluation metrics 1. Additionally, the technique supports the sampling of multiple distinct prompts that can induce similar responses from the target model 1. The paper associated with this work is dated July 31, 2026, although the source notes this may be a typographical error or a specific preprint designation 1.
评论
还没有评论,欢迎留下第一条。