Reversing the usual direction of prediction

Researchers have presented a language-model inversion method intended to recover prompts from generated responses while treating the target model as a black box. The paper, highlighted on November 19, 2026, calls the technique PTP, for previous-token prediction, and reports that it can produce near-exact prompt reconstructions without access to a target model's internal weights or probability scores.

Ordinary large language models generate a response by repeatedly predicting the next token from the text that came before it. Reconstructing the prompt from an output is harder because many different prompts can lead to similar responses. Earlier inversion work has often framed the task as recovering a semantically plausible prompt, using pretrained sequence-to-sequence systems, external datasets or privileged access to model information.

PTP instead builds a separate inverse model from scratch. Its training data is generated synthetically by querying the target model, creating prompt-and-response examples that capture that model's own behaviour. The inverse system learns to predict preceding tokens, establishing a generation process that runs in the opposite direction from conventional next-token prediction.

The authors say sampling allows the method to generate multiple candidate prompts. Those candidates may differ in wording but are selected to induce similar responses when submitted to the forward target model. That acknowledges the many-to-many nature of the problem rather than assuming every response has one uniquely recoverable origin.

Reported gains and security implications

In the paper's reported experiments, PTP outperformed prior approaches across token-based measures for both prompt and response reconstruction. The authors also report generalisation across datasets and some transferability when reconstructing prompts behind outputs produced by models other than the one used to generate training examples. Those results come from the paper and require independent replication before broader conclusions can be drawn.

The black-box setting is significant because many commercial models expose only an input-and-output interface. A technique that learns an inverse through repeated interactions could potentially be studied without direct access to proprietary parameters. At the same time, generating the synthetic training set still requires queries to the target, and the evidence supplied with the paper summary does not establish the cost, scale or real-world reliability of that process.

Prompt reconstruction has applications in model analysis, but it also raises privacy and security concerns. Outputs can sometimes reveal information about hidden instructions or user inputs, and stronger inversion methods could help measure that leakage. The same capability could be misused to probe confidential prompts. Practical deployment therefore depends not only on reconstruction accuracy but also on access controls, query monitoring and clear rules for handling sensitive data.

The work is a research proposal rather than evidence that arbitrary prompts can always be recovered. Its main contribution is a functional inversion strategy: learn the reverse relationship from the target model's own generated examples, then use previous-token prediction to search for inputs consistent with an observed output.