Autoregressive (AR) models normally write one token at a time, which is accurate but slow for long answers.
ReFusion is a new way for AI to write text faster by planning in chunks (called slots) and then filling each chunk carefully.