Diffusion Models Generate Creative Chess Puzzles with Controllable Themes

Summary

The paper presents chess puzzle generation as a demanding test of computational creativity and reasoning, since changing one piece can invalidate an otherwise complete solution. It introduces a masked diffusion model that can generate puzzles conditioned on tactical themes and partial board positions, using a non-directional diffusion process. An auxiliary task for simultaneous best-move prediction improves solution uniqueness by 11.6% and theme-conditioning accuracy by 2.5%. The authors then adapt Denoising Diffusion Policy Optimization into a reinforcement-learning framework to optimize both objectives. This training increases the yield of positions that are both unique and matched to the requested theme by 89.1%. The work also releases what it describes as the first open-weight models for chess-puzzle generation, enabling further research into controllable creative generation.