PTP uses a functional approach to invert black-box LLMs for near-exact prompt reconstruction without requiring access to model weights or logits. The method trains an explicit inverse language model from scratch using synthetic data to map outputs back to original prompts.
HOW THIS AFFECTS YOU
●
researcherYou can reconstruct prompts from outputs without needing weights or logits.
●
policyThis shows how easy it is to extract original prompts from model responses, which is a privacy risk.