Subliminal Clocks: Latent Time in Diffusion Language Models

Subliminal Clocks, an arXiv paper (arXiv:2607.01774) submitted on 2 Jul 2026, demonstrates that diffusion language models internally represent a notion of denoising time and that this signal can be extracted and manipulated. The paper, authored by Maximo Rulli and coauthors from Sapienza University of Rome, EPFL and independent researchers, reports that the latent timestep signal lives in the models' residual streams and is decodable with simple probes.

What did the authors find?

The authors show diffusion language models, unlike standard diffusion approaches, are not explicitly conditioned on a timestep yet still encode a latent representation related to the diffusion timestep in their residual streams. They report that this internal signal can be reliably extracted using probes across layers, that steering the model along a low-dimensional subspace associated with the inferred timestep systematically modulates denoising progress, and that such steering produces predictable changes in model confidence and entropy. The paper further analyses the geometry of this representation and finds structured, interpretable properties in activation space.

How did the paper demonstrate the timestep is present and useful?

The paper answers this by probing internal activations and by manipulating a low-dimensional subspace tied to the inferred timestep signal. First, probes applied across layers decode a signal the authors interpret as denoising progress; the paper states plainly that "denoising progress is decodable from internal activations." Second, the authors steer the model along the identified subspace and observe systematic modulation of the model's notion of denoising progress, with downstream effects on model confidence and entropy. Finally, they characterise the geometry of the discovered representation, arguing it has structured and interpretable properties in activation space.

Why it matters

If diffusion language models carry an implicit clock for denoising inside their residual streams, that changes how practitioners can inspect and control them. A decodable latent timestep offers a new diagnostic for internal model state and a potential control handle: the paper demonstrates steering in a low-dimensional subspace changes confidence and entropy predictably. That suggests researchers can both probe and influence generative dynamics in DLMs without altering training regimes or adding explicit timestep conditioning.

What to watch

Look for follow-ups that open-source the probes and the steering directions, and for empirical studies showing how manipulating the latent timestep affects generation quality across tasks. Also watch for peer-reviewed publication and any public code or demos linked from the arXiv entry; the paper notes an arXiv-issued DOI via DataCite is pending registration.

Details and context

The submission lists authors affiliated with Sapienza University of Rome, EPFL, and independent researchers, and marks Thomas Fontanari and Simone Petruzzi as equal contributors. The arXiv identifier for the preprint is arXiv:2607.01774 and the record includes links to PDF, HTML (experimental) and TeX source. The abstract frames the contribution relative to prior diffusion-language work by highlighting that DLMs are not explicitly conditioned on a timestep, which motivates probing whether models internally track denoising progress. The authors combine decoding probes, low-dimensional steering, and geometric analysis of activation space to support their claims.

The paper frames three concrete outcomes: a decodable latent timestep signal in residual streams, a mechanism to steer denoising progress via a low-dimensional subspace with predictable effects on confidence and entropy, and a geometric characterisation of the representation. The arXiv entry lists code and data links provisionally through common platforms, and indicates the manuscript was submitted on 2 Jul 2026.

This work adds a measurable internal quantity to discussions of how diffusion language models operate and how their internal state can be harnessed without explicit timestep conditioning. Practitioners interested in model interpretability, controllable generation, or alternative diffusion architectures should track subsequent releases from these authors for code and experimental details.