Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising

This paper formulates page-level slide personalization as an inverse planning problem rather than simple template selection or instruction following. It introduces SPIRE, which corrupts the visual structures of clean slides and trains two agents to collaboratively denoise and refine executable designs through reinforcement learning. The authors provide a proof that structural denoising is a consistent surrogate for personalization and argue that the multi-agent formulation strictly reduces policy-gradient variance. The abstract reports extensive experiments and superior performance, but does not provide benchmark names, numerical gains, or implementation details.