AI Learns Latent Design Intents for Personalized Slide Generation

Current AI agent-based methods for slide design often struggle with fine-grained, page-level personalization, typically relying on predefined templates or verbose user instructions. This limitation prevents them from capturing the subtle, latent design intents crucial for truly personalized slide layouts and themes. To address this, researchers have reframed Page-level Slide Personalization (PSP) as an inverse planning problem, aiming to learn design intent without prior knowledge of specific tools like PowerPoint or Beamer. However, optimizing this end-to-end without tool control is intractable. The proposed solution is SPIRE, a principled framework that approximates PSP. SPIRE works by intentionally corrupting the visual structures of clean slides, creating a verifiable task to "denoise" these corruptions. Two agents then collaboratively refine executable designs using reinforcement learning (RL). The research provides a proof that structural denoising consistently surrogates for PSP and that the multi-agent formulation reduces policy gradient variance in RL, demonstrating SPIRE's superior performance in extensive experiments.