AeSlides framework uses verifiable aesthetic metrics and reinforcement learning to improve LLM-based slide generation layout quality.

Researchers introduced AeSlides, a reinforcement learning framework with verifiable rewards designed to supervise aesthetic layout in slide generation. The framework uses a suite of metrics to quantify slide layout quality and employs a GRPO-based reinforcement learning method to optimize slide generation models.

Training on only 5K prompts with GLM-4.7-Flash, AeSlides improved aspect ratio compliance from 36% to 85%, reduced whitespace by 44%, element collisions by 43%, and visual imbalance by 28%. Human evaluation showed overall quality scores increased from 3.31 to 3.56 (+7.6%), outperforming reflection-based agentic approaches and matching Claude-Sonnet-4.5.

The work addresses a fundamental challenge in slide generation: the production process is text-centric while quality is governed by visual aesthetics, a gap that causes current models to produce aesthetically suboptimal layouts. The verifiable aesthetic paradigm provides an efficient, low-cost alternative to heavy visual reflection or large-scale dataset fine-tuning.