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

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

Jul 1, 2026

arXiv
Abstract

Slide design requires personalizing both deck themes and page layouts. Yet, current AI agent-based methods struggle with fine-grained, page-level design. Solely relying on prespecified templates or user verbose instructions, they fail to capture latent design intents, leaving Page-level Slide Personalization (PSP) unresolved. To close this gap, this work formulates PSP as an inverse planning problem. We propose to learn a design intent without assuming any knowledge of the specific executing tools (e.g., PowerPoint, Beamer) being used. However, relinquishing control over these tools makes the problem intractable to optimize end-to-end. To overcome this, we propose SPIRE, a principled framework to solve PSP approximately. By intentionally corrupting the visual structures of clean slides, SPIRE creates a verifiable task to denoise the corruption, whereby two agents learn to collaboratively refine executable designs via reinforcement learning (RL). We present a proof that structural denoising is a consistent surrogate for PSP, and that the multi-agent formulation strictly reduces policy gradient variance in RL. Extensive experiments demonstrate the superiority of SPIRE.

Select text to highlight · click a highlight to remove · saved in this browser only
Authors
Tianci Liu, Zihan Dong, Linjun Zhang, Haoyu Wang, jing Gao, Emre Kiciman, Ranveer Chandra, Wei-Ting Chen
Your notes (browser-local)
saved
Cross-links
No linked entities.
arXiv:2607.00407