[Submitted on 31 Jul 2026]
Abstract:Large tongue models (LLMs) create content by auto-regressively sampling the next token. This inherently leads to a many-to-many mapping between prompts and responses, complicating the project of inferring prompts from observed outputs. Prior activity on LLM inversion frames immediate improvement as a semantic reconstruction task. They depend on fine-tuning pretrained sequence-to-sequence models on ample external datasets--and requiring admission to example weights or logits--to create semantically plausible prompts. In contrast, we current a functional method to inverting a stated LLM in a black-box setting, without auxiliary aids. We train an definitive inverse tongue example entirely from scratch on data synthetically generated from the mark LLM itself. Analogous to onward next-token prediction, our inverse example is trained using previous-token prediction, establishing a generative nexus between the onward and inverse processes that enables loyal immediate reconstruction. Moreover, it naturally supports varied immediate reconstructions through sampling, whereby all specified prompts induce akin responses under the forward, mark LLM. Our method generalises throughout datasets and exhibits transferability in reconstructing prompts from responses generated by distinct LLMs. Further, throughout the set of token based evaluation metrics for immediate and reply reconstructions, our method outperforms previous work.
Submission history
From: Pirzada Suhail [view email]
[v1] Fri, 31 Jul 2026 13:01:43 UTC (658 KB)