“once data enters a Transformer, it remains recoverable.”
One of my fav quotes of this thread.
This paper put a light on the liminal space showing it’s possible to reverse the process hence it’s not that liminal anymore!
LLMs are injective and invertible.
In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space.
(1/6)