Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards (arxiv.org)

Researchers have introduced AeSlides, a novel framework that significantly enhances the aesthetic quality of slide generation using large language models (LLMs) through a reinforcement learning approach. Traditionally, slide generation processes rely heavily on text, often resulting in visually unappealing layouts due to a lack of effective supervision on aesthetic principles. AeSlides addresses this by introducing verifiable metrics that quantify layout quality, enabling the optimization of slide designs for aesthetics without the high costs associated with visual reflection or extensive datasets. The framework demonstrates impressive results, with a marked improvement in compliance with optimal aspect ratios from 36% to 85%, while simultaneously reducing whitespace by 44% and element collisions by 43%. Human evaluations also reflect a significant enhancement in perceived quality, with scores rising from 3.31 to 3.56, surpassing previous optimization methods. This pioneering approach not only showcases the potential for more efficient and scalable slide generation in alignment with human aesthetic preferences but also sets a precedent for integrating explicit aesthetic supervision in LLM-driven tasks.