source: arxiv artificial intelligence: dibs: diffusion-informed branch selection
level: research
solving sudoku puzzles often involves two main approaches: traditional heuristic solvers and deep learning solvers. heuristic solvers are complete and guarantee correct answers but can struggle with very hard puzzles, leading to long search times. deep learning solvers are fast but lack hard correctness guarantees, sometimes making mistakes. a new method called dibs combines the strengths of both by using a diffusion model to guide a symbolic solver's branch selection.
dibs keeps the symbolic solver's completeness, so it always finds a correct solution if one exists. the diffusion model acts as a guide, ranking candidate values for each empty cell based on the current partial assignment and simple consistency checks. this ranking helps the solver choose the most promising branches first, reducing the search space and speeding up the solving process. the authors provide a theoretical proof explaining why this ranking works.
experiments on the challenging royle dataset show that dibs outperforms existing methods. it solves more puzzles within a given time limit and reduces the number of search steps needed. the approach demonstrates how generative models can be integrated into symbolic reasoning systems to improve efficiency without sacrificing correctness. this hybrid method could be applied to other constraint satisfaction problems beyond sudoku.
why it matters: this shows how diffusion models can improve search efficiency in symbolic ai systems, making them faster for hard problems while keeping reliability.
source: arxiv artificial intelligence: dibs: diffusion-informed branch selection