Researchers from IIT Bombay and Adobe Research have created an inverse language model that reconstructs the original instruction from the text generated by a language model. The authors named the method "Previous-Token Prediction."
According to The Decoder's synopsis, the approach does not require access to the source model's weights and works with different models. The publication describes the reconstruction accuracy as near-perfect.
For companies using closed system instructions, this implies a potential risk of exposing such configurations. However, the available material consists of metadata and a synopsis of the publication rather than the full research text, leaving the experimental conditions and the method's applicability boundaries unclear.
editorial commentary
Why it matters
The likely consequence is increased attention to protecting closed system instructions and assessing leaks through model responses. The next observable signal will be the publication of the full study detailing experiments, metrics, and limitations. Significant uncertainty remains because currently only a metadata_only synopsis from a single source is available.