Diffusion as training curriculum: sub-250K-param solver hits 99.9% on Sudoku-Extreme
tyrell_turing · x · 2026-09-04
The Paradigms of Intelligence team (including Blake Richards) published a new paper, Diffusion as a Training Curriculum for Timestep-Free Iterative Reasoning:
- Method: add a persistent hidden state to a diffusion denoiser and remove timestep conditioning, yielding an anytime iterative solver whose accuracy keeps improving with inference depth far beyond the training rollout and backprop window
- Results: with under 250K parameters it reaches 99.90% exact solve on Sudoku-Extreme and 98.93% on Maze-Unique
- Surprising finding: progressive denoising isn't needed at inference — re-injecting fresh Gaussian noise each step (holding corruption at max) still converges to correct solutions via a single trajectory, with no parallel rollouts, candidate selection, or external verifiers
- Takeaway: diffusion's main contribution is a denoising training curriculum (ordered annealed corruption remains critical in training), not an inference-time sampling procedure
More from Research
- DeepMind's "LLM can't jump" argument: induction and deduction can't produce scientific revolutions — 0xsachi · 2026-09-04
- Free 39-Episode Control Bootcamp: The Control Theory That Runs Real Robots — lukas_m_ziegler · 2026-09-04
- UniReps Workshop in Paris opens call for papers on unified representations, due Oct 4 — ClementineDomi6 · 2026-09-04
- Utopia: open-source bitemporal knowledge graph gives RAG agents a memory of change — Shruti_0810 · 2026-09-04
- Emotion is an optimizer's control plane, not an irrational advisor — mimi10v3 · 2026-09-04
- New f-loss Cures Spectral Bias in Pixel-Space Flow Matching, Speeding Convergence — serrjoa · 2026-09-04