Story

On air. The free wire runs 15 minutes behind live. See pricing

World, crypto, finance and tech news, scored for impact.

TechHacker NewsFiled 5 h ago

PTP: Previous-Token Prediction based LLM Inversion for Near-Exact Prompt Reconstruction

Open the sourceReported by Hacker News, arxiv.org, 19 November 2026

Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.

Researchers led by Pirzada Suhail have introduced Previous-Token Prediction (PTP), a black-box method for reconstructing the prompts behind an LLM's outputs without access to model weights, logits, or external fine-tuning datasets. The approach trains a separate inverse language model from scratch on synthetic data generated by the target LLM, using previous-token prediction as the mirror image of the forward next-token process, and it supports sampling multiple plausible prompts that all elicit similar responses from the original model. The work, posted to arXiv as 2607.29378, reports gains over prior LLM inversion methods across token-based metrics for both prompt and response reconstruction, and the technique transfers across datasets and across responses produced by different LLMs. It matters because most commercial models are accessible only through APIs, so faithful prompt recovery has been difficult without internal signals. Next, watch for validation on larger proprietary models and whether the synthetic training cost of building a separate inverse model per target LLM proves practical.

Muhabbit's AI wrote this summary in its own words from Hacker News's report. It can get details wrong, so check the source.

Named in this story

arXiv, Pirzada Suhail