Researchers have developed an AI large language model (LLM) that can reconstruct missing passages in ancient Greek texts without having to know how many letters are missing. Called Apollo Restore, the system could help scholars recover information from damaged papyrus manuscripts and stone inscriptions that have remained incomplete for centuries.
The model was developed through the Decoding Antiquity initiative, led by the Austrian Academy of Sciences. In a study led by Hope McGovern, researchers tested its ability to reconstruct missing passages in ancient documents, literary manuscripts, and inscriptions.
How Apollo reconstructs ancient Greek texts
Apollo Restore builds on Mistral Small, a large language model with 24 billion parameters. Researchers trained it on approximately 600 million words of Greek from antiquity through 1900. Unlike earlier systems, Apollo Restore does not require an exact count of missing characters. It examines surviving words on either side of a damaged section and predicts what could have appeared between them. This approach, known as “fill-in-the-middle,” addresses a major challenge in studying ancient manuscripts.
Stone inscriptions often have regular letter spacing, allowing scholars to estimate how many characters have disappeared. Papyrus documents are harder to reconstruct because handwriting and spacing vary. When entire sections have broken away, determining the length of a missing passage may be impossible.
AI outperforms earlier restoration models
Researchers compared Apollo Restore with two existing systems, Ithaca, an AI system designed to reconstruct ancient Greek inscriptions, and a model developed by scholar Eric Cullhed.
In tests involving gaps of up to 10 characters, Apollo Restore included the correct reading among its 20 suggestions in 80.6% of documentary papyrus cases, 54.6% of literary papyrus cases, and 61% of stone inscription cases.
A new 24-billion-parameter model, Apollo Restore, can reconstruct missing passages in damaged papyri and stone inscriptions, even when the number of missing letters is unknown. pic.twitter.com/hdNuDwnU9M
— Tom Marvolo Riddle (@tom_riddle2025) September 23, 2026
The researchers also found that earlier evaluation methods favored short gaps. When they gave equal weight to gaps of different lengths, Apollo Restore maintained a substantial advantage on longer passages.
In a blind study involving 20 specialists, experts generally preferred its suggestions over those of earlier systems. Combining agreement with published readings and expert assessments, researchers estimated that its suggestions were at least as plausible as existing scholarly restorations in about 77% of documentary papyrus cases.
AI offers new reading of Herculaneum papyrus
The team also tested Apollo Restore on a philosophical papyrus associated with Philodemus. The scroll, known as P.Herc. 1667, was carbonized during Mount Vesuvius’ eruption in 79 A.D.
Scientists have previously used AI and advanced imaging to decipher Herculaneum scrolls that remained unreadable for nearly 2,000 years. Earlier research also identified a lost work by the Greek philosopher Philodemus in another scroll recovered from the ancient Roman town.
Apollo Restore proposed an alternative reading of a passage previously interpreted as “to do as is in our nature.” Its suggestion could instead mean “we are by nature diverse.”
Researchers said the alternative highlighted a possible problem with the published interpretation, although the proposed reading remains subject to scholarly review.
Apollo Restore also helped specialists complete restoration tasks faster in a small experiment. However, researchers acknowledged that unusual dialects, uneven training data, and passages containing multiple gaps remain challenging. They plan to release the model, training data, and evaluation code to support further research.
See all the latest news from Greece and the world at Greekreporter.com. Contact our newsroom to report an update or send your story, photos and videos. Follow GR on Google News and subscribe here to our daily email!
