AeSlides: Using Verifiable Rewards to Help AI Create More Visually Appealing Presentations

Introduction: For AI-Made Slides, Looking Good Is What Counts

Large language models (LLMs) have demonstrated strong potential in automated slide generation, yet a core contradiction has never been fully resolved — the generation process is driven by text, while the final quality is judged by visual aesthetics. This "modality gap" causes current models to frequently produce slides with cluttered layouts and misaligned formatting, falling far short of professional design standards.

Recently, a new paper published on arXiv introduced a novel framework called "AeSlides." By incorporating a verifiable rewards mechanism, it fundamentally incentivizes LLMs to attend to and optimize layout aesthetics during the slide generation process, offering a fresh technical approach to automated AI presentation generation.

The Core Problem: The Modality Gap Between Text-Driven Generation and Visual Aesthetics

Current mainstream LLM-based slide generation solutions face a structural challenge: models are essentially performing text sequence generation (outputting layout code in formats like HTML, LaTeX, or JSON), but the quality of slides ultimately depends on the rendered visual output — whether elements are neatly aligned, whitespace is reasonable, hierarchy is clear, and color coordination is harmonious.

Existing solutions typically rely on two approaches. The first is "heavy visual reflection," where multimodal models visually inspect rendered results after generation and iteratively revise them — a method with extremely high inference costs and low efficiency. The second involves supervised fine-tuning based on human-annotated data, but high-quality slide design data is scarce and annotation costs are prohibitive. Both approaches have obvious scalability bottlenecks.

Technical Approach: Aesthetic Optimization Driven by Verifiable Rewards

The core innovation of AeSlides lies in transforming the aesthetic quality of slide layouts into a set of "verifiable" quantitative metrics, using them as reward signals in reinforcement learning to directly guide LLMs toward optimizing visual outcomes during the generation phase.

"Verifiable rewards" refer to reward signals that can be automatically computed through deterministic rules or programs without requiring human judgment. In the slide context, these metrics may include:

• Element alignment: Whether the boundaries of text boxes, images, and other elements follow grid alignment rules
• Spacing consistency: Whether spacing between adjacent elements remains uniform
• Overlap detection: Whether elements obstruct one another
• Whitespace ratio: Whether blank areas on the page meet design aesthetic standards
• Hierarchical structure: Whether the visual hierarchy of titles, body text, annotations, and other elements is clear

These metrics do not require an additional visual model for subjective evaluation. Instead, they can be precisely calculated programmatically directly from the generated layout code — this is precisely what "verifiable" means. By integrating these quantified aesthetic metrics into a reward function, researchers can train LLMs using reinforcement learning (such as the GRPO policy optimization algorithm), causing the model to naturally favor producing aesthetically pleasing layouts during generation.

Technical Significance: From "Post-Hoc Patching" to "Innate Aesthetics"

The value of the AeSlides framework lies in achieving a paradigm shift — from "post-hoc patching" that relies on visual feedback after generation, to "innate aesthetics" that directly internalizes aesthetic constraints during the generation process.

This approach is highly consistent with an important recent trend in the LLM field. Since works like DeepSeek-R1, using verifiable rewards for reinforcement learning has become a key technical pathway for enhancing LLM capabilities. In mathematical reasoning and code generation tasks, the correctness of answers can be programmatically verified; AeSlides cleverly extends this idea to the visual design domain, demonstrating that aesthetic quality can likewise be decomposed into computable, verifiable structured signals.

Moreover, this method offers significant efficiency advantages. Compared to approaches requiring multiple rounds of visual reflection, AeSlides can produce high-quality layouts in a single forward generation pass, dramatically reducing computational overhead during inference and making it more suitable for real-world product deployment.

Industry Context: The AI Slide Generation Space Continues to Heat Up

In recent years, AI-assisted presentation generation has become a hot segment in the productivity tools space. Products like Gamma, Beautiful.ai, and Tome have all entered the market, and Microsoft Copilot has positioned PPT generation as one of the core scenarios for Office AI integration. However, user feedback widely indicates that while AI-generated slides are acceptable in content organization, they still lag noticeably behind professional designers in visual design quality.

The technical path explored by AeSlides targets precisely this pain point. If LLMs can possess "design intuition" at the generation stage rather than relying on post-processing to compensate for visual shortcomings, it would significantly enhance the practicality and user satisfaction of AI slide products.

Outlook: Verifiable Rewards May Reshape More Creative Generation Tasks

The research approach behind AeSlides carries broad inspirational significance. Applying verifiable reward mechanisms to creative generation tasks extends beyond slides — web design, poster layout, UI design, and even document formatting all face a similar modality gap of "text generation, visual judgment."

If aesthetic standards can be systematically quantified and programmatically verified, reinforcement learning can become a universal tool for bridging this gap. The exploration by AeSlides may be just the beginning. In the future, we can expect to see more research emerge that transforms "aesthetics" into "computable signals," driving a qualitative leap for AI in the creative design domain.