🛸 AlienLM: Alienization of Language for API-Boundary Privacy

Type a prompt in natural language. It gets translated into an alien language via a vocabulary-level token bijection, processed by an AlienLM-adapted Gemma-2-9B model, and the alien response is translated back to natural text — all without the model ever seeing your original words.

📄 Paper · 💻 Code · 🤗 Model

Demo scope: This Space runs translation and inference in one backend process. In the intended deployment, the translator and bijection remain on the trusted client, so only alienized text crosses the model API boundary.