What's this about?
Diffusion language models aren't told which denoising step they're on, yet this work finds they develop an internal "clock" — latent time-tracking that emerges without supervision.
What exactly is a "diffusion language model"?
Instead of writing left-to-right like GPT, a DLM starts with noise and refines the whole text over steps — the image-generation recipe applied to words, promising faster parallel generation.
Why is it worth noting?
34↑ — emergent internal structure in an architecture some bet will challenge autoregressive LLMs (see also DeepMind's DiffusionGemma).