Paper finds diffusion language models encode denoising progress in hidden states
Sapienza · hf · 2026-07-23
Subliminal Clocks studies whether diffusion language models encode denoising progress internally.
- The authors show that DLMs contain a latent representation related to the diffusion timestep inside residual streams
- Probes can extract this signal across layers, suggesting denoising progress is decodable from activations
- Steering the model along a low-dimensional subspace tied to the inferred timestep changes confidence and entropy in predictable ways
- The paper also analyzes the geometry of this representation and argues it is structured and interpretable
Related event: Study: Diffusion Language Models Have Subliminal Clocks(2 posts)→
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27