llm-inversion

The paper presents PTP, a functional approach to LLM inversion that trains an inverse language model from scratch using previous-token prediction on synthetic data from a target black-box LLM, enabling near-exact prompt reconstruction from responses and outperforming prior work.