Diffusion Model Solves Sokoban Puzzles Without a Solver

A new study on arXiv (2608.15958) demonstrates that a transformer-based discrete diffusion model, trained solely on tile completion, can achieve a 77.4% solvability rate on Sokoban puzzles, a PSPACE-complete problem. The model, which has no access to solvers, rewards, or solvability labels, learns to generate solvable puzzles by filling in masked cells. Notably, 94.5% of the remaining failures can be made solvable by removing a single wall, indicating that the model's local training objective implicitly captures global solvability properties. The research contrasts masked diffusion with autoregressive models, which factorize probability distributions in a fixed order, whereas diffusion hides a random subset of cells. This work, authored by researchers and posted on arXiv, highlights the emergence of complex, search-heavy properties from simple local objectives, with potential implications for procedural content generation and AI training methodologies.