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

Jul 16, 2026

Upload by Eliana

Slide design requires adapting both deck themes and page layouts, yet existing AI agent methods often fail at fine-grained, page-level personalization. Relying on fixed templates or verbose instructions cannot capture latent design intents, leaving Page-level Slide Personalization (PSP) insufficiently solved. The work formulates PSP as an inverse planning problem that learns design intent without assuming knowledge of specific presentation tools. It introduces Spire, using structural denoising with multi-agent reinforcement learning and proves the surrogate consistency and reduced policy-gradient variance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages39

Reading Time1 h 38 min

Introduction

Page-level Slide Personalization (PSP) Formulation

SPIRE: A Theoretical-Grounded Solution

Strong Performance w/ 7B-level Models

arXiv :2607 .00407v 1 [ cs .AI] 1 Jul 2026 Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising Tianci Liu 1 ,∗ , Zihan Dong2 , Linjun Zhang2 , Haoyu Wang3 , Jing Gao 1 , Emre Kıcıman4 , Ranveer Chandra4 , and Wei-Ting Chen4† 1 Purdue University 2 Rutgers University 3 University at Albany 4 Microsoft Abstract. Slide design requires personalizing both deck themes and page layouts. Yet, current AI agent-based methods struggle with finegrained, 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. 1 Introduction Slide decks are a primary medium for communicating ideas in academia and industry [13, 34, 40] . Yet, creating a high-quality deck is a design-intensive process. Depending on the presenter and the target audience, the same content can call for different visual treatments. These treatments are often personalized toa speaker’s habitual design, a lab’s visual identity, or an organization’s brand guidelines. In short, practical slide generation is inherently a personalized task. * Work done during an internship at Microsoft. † Corresponding author. 2 T. Liu et al. | <br>Fixed Templates | | <br>Long Instruction | | --- | --- | --- | PSP Fine-Tuning User Request Reinforcement Learning Contributions: Existing methods Ø Limited Quality High Human Effort SPIRE (ours) Ø High Quality Ø Learn to Infer Intents 1. Page-level Slide Personalization (PSP) Formulation 2. SPIRE: A Theoretical-Grounded Solution 3. Strong Performance w/ 7B-level Models Fig. 1: Illustration of Spire. Trained with reinforcement learning, Spire learns to infer design intents instead of using prespecified layout templates or lengthy user instructions as existing methods do [10, 53], offering both theoretical and empirical advantages. Recent advances in multi-modal large language models (MLLMs) have enabled agentic pipelines that automate slide generation for practical use [9, 10, 26, 30, 38, 45, 53] . These systems typically decompose slide creation into modular stages such as outlining, asset extraction, layout arrangement, and iterative refinement, thereby improving deck-level coherence via template selection, visual feedback, and multi-agent coordination [15, 16, 30, 42, 50, 53] . However, they struggle with fine-grained page-level design, which requires deciding visual hierarchy, element alignment, spacing, and styling choices. Nonetheless, existing agentic systems handle page-level layout and styling passively: Once the content is specified, the page-level design is largely overlooked, either by following generic system-defined templates [15,24,42,53], or by requiring lengthy, explicit user instructions [10,30], instead of inferring the user’s latent intent. Consequently, they are too coarse to achieve satisfactory Page-level Slide Personalization (PSP) . To close this gap, we introduce the task of agentic PSP; the concept is depicted in Fig 1 . Our for

FAQs

Why do current agentic slide generation methods struggle with page-level personalization?

They typically follow generic system templates or require lengthy explicit user instructions, which overlook fine-grained page design choices such as hierarchy, alignment, spacing, and styling needed for true PSP.

How does the proposed method formulate PSP?

It formulates PSP as a probabilistic inverse planning problem where design intent is modeled as a latent random variable inferred from user-provided page assets and a small set of reference slides.

What is Spire and how does it address the intractability of end-to-end optimization?

Spire approximates PSP by intentionally corrupting the visual structures of clean slides, then training agents to denoise the corruption via reinforcement learning so executable designs can be refined collaboratively.

Jul 16, 2026

Upload by Eliana