Researchers have developed a novel method for reconstructing prompts used to generate text from large language models (LLMs). Unlike previous approaches that required fine-tuning models or access to their weights, this new technique operates in a black-box setting. It involves training an inverse language model from scratch using synthetically generated data from the target LLM, employing a previous-token prediction strategy that mirrors the forward generation process. This method not only achieves near-exact prompt reconstruction but also supports diverse prompt generation and demonstrates transferability across different LLMs and datasets, outperforming existing methods on various evaluation metrics. AI
IMPACT Enables better understanding and potential security analysis of LLM outputs by allowing reconstruction of the input prompts.
RANK_REASON Academic paper detailing a new method for LLM prompt reconstruction. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
