AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards

Technical Summary: AeSlides

AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards

1. Problem Statement

Current Large Language Model (LLM) based slide generation suffers from a fundamental modality gap. While the generation process is text-centric (producing structured markup like HTML), the quality of a slide is judged by its visual aesthetics. This gap leads to several recurring layout deficiencies:

• Distorted Aspect Ratio: Content does not fit the standard 16:9 ratio.
• Excessive Whitespace: Inefficient use of space, leading to a sparse appearance.
• Element Collision: Overlapping elements, overflow from containers, or content falling outside slide boundaries.
• Visual Imbalance: Uneven distribution of visual weight, causing perceptual bias.

Existing solutions—such as agentic reflection (iterative refinement via VLMs) or SFT (fine-tuning on large datasets)—are either computationally expensive with limited gains or provide weak, indirect aesthetic supervision because they rely on implicit textual patterns rather than explicit visual principles.

2. Methodology

The authors propose AeSlides, a reinforcement learning (RL) framework that uses verifiable rewards to directly optimize for aesthetic layout.

A. Verifiable Aesthetic Metrics
Instead of relying on unpredictable Vision-Language Models (VLMs) as reward providers, AeSlides introduces a suite of programmatic, rule-based metrics:

• Aspect Ratio Detection: Uses a dynamic rendering procedure (via Playwright) to measure the actual rendered height/width of the content.
• Whitespace Detection: Employs a lightweight vision-based method using a local variance map. It identifies regions with low pixel variation (low-frequency regions) to quantify empty space.
• Collision Detection: Analyzes the DOM tree and element bounding boxes to identify overlapping "visual units" or boundary violations.
• Imbalance Detection: Calculates the visual centroid of the layout and measures its offset from the canvas center using an ellipse-normalized distance.

B. Reinforcement Learning Framework
The framework utilizes Group Relative Policy Optimization (GRPO) to optimize the model. Key technical enhancements include:

• Reward Shaping: To prevent optimization instability, the authors use asymmetric quadratic rewards for aspect ratios and smoothstep transformations for monotonic metrics.
• Reward-Decoupled Normalization (GDPO): To prevent "reward collapse" (where a single high-variance reward dominates the others), each reward component is normalized independently within a group before being aggregated.
• KL Divergence Regularization: Retained to ensure the model maintains semantic fidelity and does not deviate too far from the high-quality content-generation capabilities learned during SFT.

3. Key Contributions

1. New Paradigm: Introduces the concept of "verifiable rewards" for multimedia generation, bridging the modality gap between text generation and visual evaluation.
2. Metric Suite: Develops a highly accurate, low-latency, and zero-cost (compared to VLM APIs) suite of metrics for slide layout assessment.
3. AeSlides-Reward-Bench: A new dataset with fine-grained annotations across 21 aesthetic subcategories to facilitate meta-evaluation.
4. Robust RL Implementation: A specialized training pipeline that addresses common RL issues like reward hacking and signal collapse in multi-reward settings.

4. Results

The experiments conducted on the GLM-4.7-Flash model demonstrate significant improvements:

• Quantitative Gains: Using only 5k training prompts, AeSlides improved aspect ratio compliance from 36% to 85%, reduced whitespace by 44%, decreased collisions by 43%, and mitigated visual imbalance by 28%.
• Comparative Performance: AeSlides outperformed agentic reflection methods (like DeepPresenter) and model-based reward optimization. Notably, it even edged out the proprietary Claude-Sonnet-4.5 in human evaluations.
• Human Evaluation: Scores increased from 3.31 to 3.56 (+7.6%), confirming that the programmatic metrics are well-aligned with human aesthetic preferences.
• VLM Critique: The paper provides a critical analysis showing that current frontier VLMs (like GPT-5.2) are unreliable evaluators for slide layouts, often suffering from "perceptual blind spots" and preference bias.

5. Significance

AeSlides demonstrates that for tasks where the output is visual but the generation is textual, explicit programmatic supervision is superior to implicit learning or expensive multimodal feedback. This work provides a scalable and efficient blueprint for aligning LLMs with complex, multi-dimensional aesthetic standards in various multimedia creation tasks (e.g., web design, UI generation, and graphic design).