Conditional chess-puzzle generation with masked diffusion models (arXiv:2609.38577v1)
The paper proposes a masked, non-directional diffusion approach for conditional generation of chess puzzles that can be conditioned on tactical themes and partial board positions. It introduces a simultaneous best-move prediction auxiliary task that improves solution uniqueness by 11.6% and theme-conditioning accuracy by 2.5%, applies an RL adaptation of Denoising Diffusion Policy Optimization (DDPO) to boost yield of unique, theme-matching positions by 89.1%, and releases open weights for the models (Appendix B).