Search papers, labs, and topics across Lattice.
This paper investigates how Diffusion Language Models (DLMs) internally represent denoising progress despite not being explicitly conditioned on timesteps. By extracting a latent representation related to the diffusion timestep from the model's residual streams, the authors demonstrate that this information can be manipulated to influence model confidence and entropy. The findings reveal structured properties in the activation space, enhancing our understanding of how DLMs process temporal information during generation tasks.
DLMs encode latent representations of denoising progress that can be extracted and manipulated, revealing a surprising level of internal structure.
Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive models. Unlike standard diffusion-based approaches, DLMs are not explicitly conditioned on a timestep, raising a natural question: do these models internally represent denoising progress, and how is such information used downstream? In this work, we show that DLMs do in fact encode a latent representation related to the diffusion timestep within their residual streams. We find that this signal can be reliably extracted using probes across layers, indicating that denoising progress is decodable from internal activations. We further demonstrate that steering the model along a low-dimensional subspace associated with the inferred timestep allows us to systematically modulate its notion of denoising progress, leading to predictable changes in model confidence and entropy. Finally, we analyse the geometry of the identified representation, showing that it exhibits structured and interpretable properties in activation space, and shedding light on how such a signal is processed by these models.
