Masked diffusion language models LLaDA-1.5 and Dream-7B, which are not explicitly conditioned on a timestep, carry a denoising-progress (latent timestep) signal in their residual streams whose per-timestep mean vectors concentrate in a low-dimensional subspace (most variance in fewer than three principal components) tracing an ordered, parabola-like curve

- Rulli et al. probe the residual streams of LLaDA-1.5 and Dream-7B and find a latent signal related to the diffusion timestep (fraction of tokens still masked) that is decodable across layers; the per-timestep mean vectors have most of their variance in fewer than three principal components (a single PC captures a large fraction, "heavily concentrated above 90% explained variance" intra-layer), and their top-2 PC projection follows a shared parabola-like trajectory across layers. [rulli-etal-2026-subliminal-clocks] - Steering restricted to the span of the top principal components (e.g. k=2 at layer 29 of LLaDA) resembles unrestricted mean-vector steering while the orthogonal remainder produces incoherent effects, which is evidence that the subspace is a causally effective direction set rather than proof of a natural geometry; Dream shows a larger endpoint spread across layers and outlying first/last steps, so the geometry is model-dependent. [rulli-etal-2026-subliminal-clocks]