What's this about?
Small continuous register states carry progress after earlier generated chunks are cleared; authors report gains up to 8.5 points on math and 19.5 on code versus text carry (3 HF upvotes).
What exactly is register token?
A fixed-position hidden-state slot trained to retain information across chunks. Think of a compact scratchpad, not an indefinitely faithful memory of every prior word.
What's the community saying?
3 HF upvotes; author-reported preprint.
- Significance: bounded memory may reduce growing-context overhead.
- Significance: results apply to tested diffusion-model settings.
- Significance: information loss and broader transfer remain to test.