EVERY SIGNAL, SORTED.

Germany, Google, Ulm University, Self-Training, Language Models, Vision-Language Models, Synthetic Data, Model Evaluation, Curriculum Design and AI Reliability

A model wrote its own curriculum. Test what the synthetic teacher forgot

On-Policy Self-Training lets a model invent problems, privileged reasoning and token-level lessons from broad topic names. A self-consistent classroom still needs an exam written from outside the room.