Aeneas is the first AI model designed to restore, date, and contextualize ancient Latin inscriptions, revolutionizing historical research by retrieving textual parallels and decoding damaged Roman texts with high accuracy through multimodal input and open-access tools. (Source: Image by RR)

Transformer-Based System Bridges Visual and Textual Clues to Complete Ancient Messages

Researchers have introduced Aeneas, the first AI model designed specifically to contextualize ancient Latin inscriptions, offering historians a powerful new tool to restore, date and attribute fragmentary texts. Inscriptions from the Roman world have long served as critical windows into daily life, politics, and culture, but many are damaged or incomplete. Traditionally, historians spend years manually searching for similar inscriptions to piece together meanings and chronology. Aeneas accelerates this process by scanning over 176,000 inscriptions and retrieving parallels in seconds, using advanced neural networks to match content and historical context.

Aeneas, as noted in deepmind.google, builds upon a previous model, Ithaca, which was developed to analyze ancient Greek texts. Unlike its predecessor, Aeneas uses a multimodal approach, combining both textual and visual data—such as photographs of inscriptions—to boost accuracy in restoration and attribution. Notably, Aeneas can handle inscriptions with unknown-length gaps, a challenge for most models, and can assign inscriptions to one of 62 ancient Roman provinces with 72% accuracy. It also dates texts within a 13-year window on average, offering historians quantifiable insights where uncertainty previously dominated.

The model uses a transformer-based decoder and embedding techniques to create a “historical fingerprint” for each inscription. It was trained using data harmonized from leading epigraphic databases like EDR, EDH, and EDCS-ELT. In real-world tests, Aeneas has shown Top-20 restoration accuracy of 73% for ten-character gaps, and 58% for unknown-length gaps—high benchmarks for such tasks. When tested on the Res Gestae Divi Augusti, a historically contested Roman text, Aeneas produced a probabilistic dating that echoed both major scholarly hypotheses, demonstrating how the model can meaningfully participate in long-standing historical debates.

To ensure accessibility and foster collaboration, the team has made Aeneas freely available through predictingthepast.com and open-sourced its code and dataset. Aeneas also supports a new digital syllabus to help educators integrate AI into historical studies. In user trials with 23 professional historians, the model significantly enhanced researchers’ confidence and discovery of new inscriptional parallels. By blending human expertise with AI contextualization, Aeneas represents a transformative step forward in both classical scholarship and AI-assisted research.

read more at deepmind.google