Pinned post
Deployed LLMs and users generate millions of conversations every day.
These are full of useful learning signals, yet we don't use them for training.
We introduce self-distillation for learning directly from user conversations – no rewards, no labels, no extra models.
These are full of useful learning signals, yet we don't use them for training.
We introduce self-distillation for learning directly from user conversations – no rewards, no labels, no extra models.
