Benchmark for evaluating AI slide generation tackles cross-system comparability challenge

Researchers have unveiled SlidesGen-Bench, a standardized evaluation framework designed to assess how well different AI systems generate presentation slides. The benchmark addresses a persistent challenge: with a growing array of slide generation tools—some built on code, others on image synthesis—there's been no unified way to compare their output fairly.

The core problem#

As large language models have become more capable, they've spawned various approaches to automating slide creation. Yet evaluating these diverse systems has remained difficult. Existing assessment methods either produce inconsistent scores across different architectures or depend on subjective human judgment that lacks reproducibility.

How the benchmark works#

Rather than demanding a single generation method, SlidesGen-Bench grounds evaluation in the visual domain: it treats all slide outputs as rendered images, making the framework agnostic to whether slides were generated through code, templates, or image-based synthesis.

The framework assesses slides across three quantifiable dimensions:

• Content: How well the slide communicates its intended information
• Aesthetics: Visual design quality and presentation
• Editability: How easily generated slides can be modified or refined

Alignment with human preferences#

To ensure the metrics actually reflect what users prefer, the team built Slides-Align1.5k, a dataset of human preference judgments covering slides from nine mainstream generation systems across seven different scenarios. Testing showed SlidesGen-Bench achieves stronger correlation with human judgment than prior evaluation pipelines.

The researchers have made their code and data publicly available to support further development in this area.