Andrii Bidochko

• Updated: August 24, 2026
• 7 min read

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

Direct Answer

The paper introduces SPIRE (Structural Planning via Inverse REasoning), a novel framework that treats page‑level slide personalization as an inverse planning problem and learns latent design intents without hard‑coding any specific presentation tool. By corrupting slide structures and training two collaborative agents to denoise them, SPIRE can generate highly personalized decks that respect both global themes and fine‑grained layout preferences, a capability that has been missing from existing AI‑driven slide designers.

Background: Why This Problem Is Hard

Creating a presentation deck is more than picking a template; it involves aligning visual hierarchy, typography, color palettes, and layout constraints with a user’s implicit design intent. Traditional AI agents either:

• Rely on pre‑specified templates that limit creativity and cannot adapt to nuanced user preferences.
• Consume verbose textual instructions, which are brittle because designers rarely articulate every layout decision.
• Operate on a single‑level abstraction, handling deck‑wide themes but ignoring page‑level variations such as image placement or bullet‑point alignment.

These approaches fail to capture the latent, often subconscious, design intents that professional designers embed in each slide. Moreover, the execution environment (PowerPoint, Google Slides, LaTeX Beamer, etc.) introduces tool‑specific constraints that are difficult to model directly. The combination of hidden intent, fine‑grained layout requirements, and tool heterogeneity makes page‑level slide personalization (PSP) an open research bottleneck.

What the Researchers Propose

SPIRE reframes PSP as an inverse planning problem: instead of directly commanding a slide editor, the system infers the underlying design intent that would have produced a given clean slide. The framework consists of three high‑level components:

1. Structural Corruption Engine: Takes a pristine slide and deliberately distorts its visual structure (e.g., shuffling elements, altering alignments) to create a “noisy” version.
2. Design Intent Agent (DI‑Agent): Learns a latent representation of the user’s design intent by observing the relationship between corrupted and clean slides.
3. Execution Agent (EX‑Agent): Uses the inferred intent to iteratively refine a draft slide until it matches the clean target, operating purely through reinforcement learning signals.

The key insight is that denoising corrupted slides is a tractable surrogate for the original inverse planning objective. By training the two agents jointly, SPIRE sidesteps the need to model the idiosyncrasies of any specific presentation tool.

How It Works in Practice

Conceptual Workflow

The end‑to‑end pipeline can be visualized as a loop:

1. Input Collection: A user provides a high‑level theme (e.g., “modern tech conference”) and optionally a few exemplar slides.
2. Corruption Phase: The Structural Corruption Engine generates a set of noisy slides by perturbing layout, typography, and visual hierarchy.
3. Intent Inference: The DI‑Agent processes the noisy‑clean pairs, extracting a compact latent vector that encodes the user’s implicit design preferences.
4. Iterative Refinement: The EX‑Agent receives the latent intent and a draft slide, then takes actions (move element, resize, recolor) guided by a reward function that measures structural similarity to the clean target.
5. Convergence Check: When the reward exceeds a predefined threshold, the refined slide is emitted as the final personalized output.

Agent Interaction Details

• Collaborative Learning: Both agents share the same replay buffer of corrupted‑clean pairs, allowing the EX‑Agent to benefit from the DI‑Agent’s evolving intent embeddings.
• Policy Gradient Variance Reduction: The paper proves that the multi‑agent formulation yields a lower‑variance gradient estimator, accelerating convergence in reinforcement learning.
• Tool‑Agnostic Execution: Because the EX‑Agent’s actions are defined in an abstract “design space” (e.g., move element to (x, y)), a thin adapter can translate them to PowerPoint, Google Slides, or LaTeX commands without retraining the core model.

What Sets SPIRE Apart

Unlike template‑driven generators, SPIRE does not require a fixed library of slide layouts. Instead, it learns a continuous design intent space that can be sampled to produce infinitely many variations. Moreover, the structural denoising objective provides a verifiable training signal, eliminating the need for costly human‑in‑the‑loop annotations.

Evaluation & Results

Experimental Setup

The authors benchmarked SPIRE against three baselines:

• Template‑Only Generator: Selects the best matching pre‑designed template based on keyword similarity.
• Instruction‑Based LLM: Uses a large language model to generate slide markup from natural‑language prompts.
• End‑to‑End RL Agent: Trains a single RL agent to map user intent directly to slide actions, without the denoising surrogate.

Evaluation metrics included:

• Structural Fidelity: Pixel‑level and layout‑level similarity to ground‑truth personalized slides.
• Design Intent Alignment: Cosine similarity between inferred intent vectors and human‑annotated intent embeddings.
• User Preference Score: A blind study where 50 professional designers rated the aesthetic appeal and functional suitability of generated decks.

Key Findings

SPIRE consistently outperformed all baselines:

• Structural fidelity improved by 23 % over the template‑only approach and by 15 % over the instruction‑based LLM.
• Design intent alignment showed a 0.42 increase in cosine similarity compared to the single‑agent RL baseline.
• In the user study, SPIRE‑generated decks received an average rating of **4.6/5**, versus **3.8/5** for the best competing method.

Beyond raw numbers, qualitative analysis revealed that SPIRE could adapt a single global theme to produce slides with distinct visual hierarchies—e.g., emphasizing data visualizations on one page while prioritizing textual storytelling on another—mirroring how human designers tailor each page.

Why This Matters for AI Systems and Agents

SPIRE’s contributions ripple across several domains of AI‑driven productivity:

• Agent‑Centric Design: By framing personalization as inverse planning, SPIRE offers a blueprint for building agents that infer hidden user goals from observable outcomes—a pattern applicable to code generation, UI layout, and even robotic manipulation.
• Tool‑Agnostic Orchestration: The abstract action space enables seamless integration with existing presentation platforms via lightweight adapters, reducing engineering overhead for enterprises that rely on PowerPoint or Google Slides.
• Reduced Human Annotation Burden: Structural denoising supplies a self‑supervised training signal, allowing developers to scale personalization models without costly labeling pipelines.
• Enhanced Workflow Automation: Companies can embed SPIRE into Workflow automation studio to automatically generate meeting decks from meeting transcripts, aligning visual style with corporate branding without manual tweaking.
• Competitive Edge for AI Marketing Agents: Marketing teams can leverage SPIRE to produce campaign‑specific slide decks that reflect brand voice and visual identity, accelerating go‑to‑market cycles. See AI marketing agents for related use cases.

What Comes Next

While SPIRE marks a significant step forward, several avenues remain open:

• Cross‑Modal Intent Learning: Extending the latent intent space to incorporate audio (e.g., speaker tone) or video cues could enable richer multimodal presentations.
• Real‑Time Interaction: Integrating a feedback loop where designers can correct or steer the EX‑Agent mid‑generation would blend human creativity with AI efficiency.
• Scalability to Large Decks: Current experiments focus on decks of up to 15 slides; optimizing the reinforcement learning horizon for longer presentations is an engineering challenge.
• Domain‑Specific Adaptation: Fine‑tuning SPIRE for specialized sectors—such as scientific conferences (Beamer) or financial reporting (Excel‑linked slides)—could unlock niche markets.
• Open‑Source Tooling: Publishing a modular SDK would allow the community to build custom adapters for emerging presentation platforms, fostering an ecosystem around inverse‑planning agents.

Developers interested in experimenting with SPIRE can start by exploring the UBOS platform overview, which provides a sandbox for multi‑agent reinforcement learning and structural denoising pipelines.

References

For a complete technical description, see the original preprint: Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.