Self-Regularized Flow Language Models
Self-supervised training for continuous flow language models that reuses the denoising trajectory as its own signal.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv