Firehose

This paper investigates whether diffusion language models can continue reasoning across generation chunks without keeping earlier text in context, and whether using a fixed-size "register" can improve performance. Practitioners might care about this because it could lead to more efficient and flexible language generation models.