PTP reconstructs LLM prompts in a black-box setting
Reported by Hacker News and ecosistemastartup2 sources ↓
PTP is a method for reconstructing prompts from an LLM’s responses without access to its weights or logits. Its authors say it trains an inverse model on synthetic data generated by the target LLM.
Key points
- PTP trains a model to predict previous tokens, reversing an LLM’s usual next-token process
- The inverse model uses synthetic data from the target LLM, with no external datasets or auxiliary models
- The paper reports better token-based evaluation results than prior methods and transfer to outputs from other LLMs
- The method does not guarantee identical prompt recovery; it can generate multiple candidate reconstructions
Sources
Details
Ecosistema Startup reports that PTP, described in paper arXiv 2607.29378, reconstructs prompts from model responses using only the ability to query the target LLM. The method does not require access to the target’s weights or logits.
Rather than train a model to infer a prompt’s meaning, PTP trains an inverse model from scratch to predict the preceding token. Training data is generated by the target LLM itself, and the inverse model can sample multiple possible prompt reconstructions.
The paper reports that PTP generalizes across datasets and transfers to responses from LLMs other than the one used to train the inverse model. It also reports better token-based evaluation results than prior work. The article describes the reconstructions as near-exact, not guaranteed identical.
Written from coverage as of
Images
- ecosistemastartup