Recurrent denoisers: adding a persistent hidden state makes diffusion an anytime solver

burkov · x · 2026-09-11

The paper explores iterative solvers that improve simply by running longer at inference. While standard diffusion carries information only in its noisy state, the authors test whether adding a persistent hidden state to a timestep-free diffusion denoiser yields an anytime solver that refines solutions without explicit time conditioning or fixed rollout lengths.

Recurrent denoisers — looped Transformers for Sudoku-Extreme and local convolutional blocks for Maze-Unique — were trained on millions of unique-solution puzzles using 4-step truncated back-prop segments and an annealed noise schedule. At inference, the same shared update runs for thousands of steps with the hidden state carried forward, so accuracy grows with compute.

Original post →

More from Research

Research channel →