|
Previous-Token Prediction Based LLM Inversion for Prompt Reconstruction
Researchers have developed a black-box method called PTP that reconstructs original prompts from LLM-generated responses by training an inverse language model using previous-token prediction on synthetically generated data from the target LLM. Unlike prior approaches that require model weights, logits, or external datasets, this functional method works without auxiliary aids and can generate multiple diverse prompts that produce similar responses, while outperforming existing techniques on token-based evaluation metrics.
Read Full Article →
← Back to latest