Subliminal Clocks: Latent Time Modelling in Diffusion Language Models
The study found that Diffusion Language Models (DLMs) internally encode information about denoising progress, represented by a latent variable related to the diffusion timestep. Researchers trained probes to predict this progress using residual streams, confirming consistent information representation across network layers. Explicitly extracting and steering this latent signal successfully steered model behavior in interpretable ways.