SPIRE: Inverse Planning for Agentic Slide Generation, ECCV 2026

Tianci Liu, Zihan Dong, Linjun Zhang, Haoyu Wang, jing Gao, Emre Kiciman, Ranveer Chandra and Wei-Ting Chen submitted the paper "Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising" to arXiv (arXiv:2607.00407) on 1 Jul 2026, with the submission file listed as 8,725 KB. The authors introduce SPIRE, a framework that treats Page-level Slide Personalization (PSP) as an inverse planning problem and trains collaborative agents to recover executable, personalized slide designs from deliberately corrupted visuals.

How does SPIRE work?

SPIRE formulates PSP as inverse planning and trains two agents to denoise intentionally corrupted slide structure via reinforcement learning (RL), producing executable designs without assuming knowledge of the target authoring tools. The system corrupts the visual structures of clean slides to create a verifiable denoising task; two agents learn to collaboratively refine those corrupted layouts into designs that can be executed in tools such as PowerPoint or Beamer, even though SPIRE itself does not assume knowledge of those tools.

The framework centers on learning latent design intents rather than relying on prespecified templates or long user instructions. That design-intent representation is learned by treating the generation process as reversing a corruption, so the agents optimize policies that map corrupted inputs back to coherent, personalized slide pages.

How did the authors validate SPIRE?

The paper presents both theoretical and experimental validation: the authors give a proof that structural denoising is a consistent surrogate objective for Page-level Slide Personalization, and they show that the multi-agent formulation strictly reduces policy gradient variance in RL. The abstract summarizes the empirical side as "Extensive experiments demonstrate the superiority of SPIRE." The arXiv entry also lists the paper with Comments: ECCV 2026 and provides an arXiv DOI (https://doi.org/10.48550/arXiv.2607.00407).

The evaluation approach therefore combines a formal argument about the surrogate loss with empirical experiments. The multi-agent design is singled out as reducing variance in policy gradients, a concrete training benefit cited in the paper.

Why it matters

SPIRE directly targets a gap current agent-based slide tools leave open: fine-grained, page-level personalization. Existing methods either apply fixed templates or ask users for verbose instructions and so fail to capture the latent intents that drive layout and theme choices. By separating design intent from the specifics of execution tools and learning to denoise structural corruptions, SPIRE gives agents a pathway to infer those latent intents automatically.

That separation matters for two reasons. First, a tool-agnostic intent representation lets downstream rendering systems or human designers reuse learned intents across PowerPoint, Beamer or other formats. Second, the multi-agent RL setup that reduces policy gradient variance makes training more stable, which matters for deploying agentic design systems at scale where brittle policies would otherwise hinder consistent page-level outputs.

What to watch

Watch for the ECCV 2026 presentation and for the code, demos and dataset links the arXiv record points toward (the entry includes toggles for Code, Data and Media and mentions services such as Replicate and Hugging Face). Also track follow-up evaluations that quantify how much the multi-agent formulation reduces variance and how those gains translate into perceptible improvements in page-level personalization compared with template- or instruction-driven baselines.

Paper reference details: arXiv:2607.00407 [cs.AI], submitted Wed, 1 Jul 2026 04:05:47 UTC (8,725 KB), DOI 10.48550/arXiv.2607.00407. The author list is Tianci Liu, Zihan Dong, Linjun Zhang, Haoyu Wang, jing Gao, Emre Kiciman, Ranveer Chandra and Wei-Ting Chen. The arXiv entry includes links to PDF, TeX source and several code/data toggles.