PosterSum: A multimodal benchmark for scientific poster summarization

Scientific posters pack dense visuals and terse text into a single canvas, which makes summarization harder than standard document settings. PosterSum targets this gap with a multimodal benchmark for generating paper abstracts from poster images.

The dataset includes 16,305 posters covering multiple fields, and the evaluation compares end-to-end approaches with segment-then-summarize pipelines. The results highlight that layout-aware segmentation is still critical for strong performance.

PosterSum provides a practical testbed for multimodal summarization systems that need to reason over images, text blocks, and layout structure simultaneously.

Paper: https://aclanthology.org/2025.findings-ijcnlp.114/