Diffusion as a Training Curriculum for Timestep-Free Iterative Reasoning
The paper reframes diffusion as a training method for iterative solvers, not an inference-time sampling recipe.
The authors add persistent hidden state to a diffusion denoiser and remove timestep conditioning, so one shared update can run for arbitrary depth. Accuracy continues improving beyond the training rollout, reaching 99.90% exact solve on Sudoku-Extreme and 98.93% on Maze-Unique. At inference, the model still works when non-clue variables are repeatedly replaced with fresh maximum corruption noise, suggesting progressive denoising is not required then. The training process still depends on ordered annealed corruption, which the authors argue acts as the useful curriculum. ArXiv · AI/CL/LG's note
The authors add persistent hidden state to a diffusion denoiser and remove timestep conditioning, so one shared update can run for arbitrary depth. Accuracy continues improving beyond the training rollout, reaching 99.90% exact solve on Sudoku-Extreme and 98.93% on Maze-Unique. At inference, the model still works when non-clue variables are repeatedly replaced with fresh maximum corruption noise, suggesting progressive denoising is not required then. The training process still depends on ordered annealed corruption, which the authors argue acts as the useful curriculum. ArXiv · AI/CL/LG's note
score 6