Diffusion as a Training Curriculum for Timestep-Free Iterative Reasoning (arxiv.org)

Researchers have introduced an innovative approach to combining diffusion models and recursive reasoning to create an "anytime solver" capable of solving complex problems, such as Sudoku-Extreme and Maze-Unique, with remarkable accuracy rates of 99.90% and 98.93%, respectively. By integrating a persistent hidden state in a diffusion denoiser while eliminating timestep conditioning, the model allows for iterative reasoning without the constraints of traditional rollouts or parallel candidate selection. This approach enables the solver to improve its accuracy continuously with deeper inference while maintaining near-perfect performance even with maximum corruption injected during inference. This development is significant for the AI/ML community as it redefines how information is carried across iterations in reasoning models, highlighting the importance of the training curriculum rather than the sampling method used in inference. The findings suggest that a simple noise-injection mechanism can efficiently navigate the solution landscape, presenting a potential shift in methodology for designing future AI systems that require iterative reasoning. By showcasing that ordered annealed corruption is critical during training, this research paves the way for more effective training practices that could enhance the capabilities of various AI applications.