Aeneas transforms how historians connect the past

Introduces Aeneas, a multimodal generative neural network for contextualizing ancient inscriptions. It uses a transformer-based decoder and specialized networks for text restoration, dating, and geographical attribution using images. The model retrieves parallels using embeddings, encoding textual and contextual information into a historical fingerprint. It achieves state-of-the-art performance in restoring damaged texts (73% Top-20 accuracy for known length gaps, 58% for unknown length) and dating (within 13 years of historian estimates). It also demonstrates accurate geographical attribution to ancient Roman provinces (72% accuracy). The post details a collaborative study with historians showing Aeneas accelerates work and increases confidence in epigraphic tasks.