Subliminal Clocks: Latent Time Modelling in Diffusion Language Models
M. Rulli, T. Fontanari*, S. Petruzzi*, F. Alvetreti, G. Strano, D. Crisostomi, G. Nikolaou, T. Mencattini, A. Santilli, E. Rodolà, S. Scardapane, A. Devoto
2026
We show that Diffusion Language Models internally encode denoising progress, and that this signal can be probed, steered, and geometrically characterised.

