Rep2Text: decoding text from one hidden representation
A research paper introduces Rep2Text, a method for reconstructing text from a single hidden representation taken at the last token. Its approach combines a trainable adapter with a decoder language model. The paper reports that for short texts of about 16 words, it can recover roughly half of the original words while preserving overall meaning. Exact wording becomes harder to recover as input length increases, though the main ideas may remain recoverable.
The result suggests that a hidden representation can retain information about its source text. The authors identify potential privacy risks, but the abstract does not demonstrate an attack on a deployed service or establish how often such recovery would work in operational conditions. This is a research finding about representation leakage, not a measurement of product-level exposure.