AI & Archaeology · Ancient Greek texts; scroll buried AD 79 · Herculaneum (PHerc. 1667), Campania, Italy
Meet Apollo Restore, an AI that fills gaps in damaged ancient Greek texts
In a blind test, 20 specialists rated its restorations at least as good as human ones in 77% of cases, but the study is not yet peer reviewed.
Not yet peer reviewed
Herculaneum (PHerc. 1667), Italy A new artificial intelligence model suggests the missing letters in damaged ancient Greek texts. It is called Apollo Restore. Its makers report that it beats the best earlier published models on papyri and stone inscriptions. It also improved the published reading of a papyrus roll carbonised when Vesuvius erupted in AD 79.
The team describes the model in a preprint posted on arXiv. A preprint has not yet passed peer review. Apollo Restore comes from the Decoding Antiquity initiative, which builds specialised language models for historical languages and manuscripts. The Austrian Academy of Sciences leads the initiative.
How the model fills a gap
Ancient texts on papyrus and stone are often broken. Scholars call the missing stretches lacunae: physical gaps where letters have been lost. Papyrologists and epigraphists propose restorations for them from their knowledge of the language and of similar texts.
Apollo Restore is a large language model with 24 billion parameters. The team built it by fine-tuning an existing model, Mistral Small, to fill in the middle. It learns to write the missing span between the surviving text before and after a gap. It does not need to be told in advance how many letters are missing.
The authors say that, to their knowledge, it is the first large-scale model of its kind for historical Greek. They add that it is the first for any ancient Mediterranean language.
Tested against earlier models and human experts
The team first tested the model the way earlier studies did, on short gaps of up to ten characters. The model gives a ranked list of possible restorations. The test asked whether the correct one appeared in its top twenty.
It did for 80.6% of gaps in documentary papyri, 54.6% in literary papyri and 61.0% in stone inscriptions. The authors report that these scores are 1.6, 2.6 and 1.4 times those of the strongest published models on the same three groups.
The team argues that the older test inflates scores. It is biased towards trivially short gaps. So they also scored the model with a measure balanced for gap length. On that measure, its lead over the best published models grows to 2.3, 3.5 and 1.6 times.
They also gave the model wrong hints about how long a gap was. They report that its performance fell off only gradually.
In a blind study, 20 expert papyrologists, epigraphists and philologists compared its output with that of the strongest baseline model. They strongly preferred Apollo Restore. They judged its performance at least as good as human restorations in 77% of cases.
The model was also tried on PHerc. 1667, one of the carbonised papyrus rolls from Herculaneum, the Roman town buried by Vesuvius. The roll was digitally unrolled and edited only after the model's training data had been compiled. That edition was therefore not among the material the model learned from. The authors report that Apollo Restore improves its published reading.
The abstract leaves some questions open. It does not say which passages of PHerc. 1667 changed, or how. In the other 23% of the blind comparisons, the experts did not rate the model as good as human restorations.
For literary papyri, the correct answer was missing from the top twenty in almost half of the test gaps. The main figures cover gaps of ten characters or fewer. The model returns a ranked list of up to twenty candidates for each gap.
Sources
Paper: Apollo Restore: A Foundation LLM for Historical Greek Optimized for Fill-in-the-Middle Restoration of Ancient Greek Texts. arXiv preprint arXiv:2609.22455. DOI: 10.48550/arXiv.2609.22455
Spotted a mistake? Tell the editor about a spelling error or a fact that is wrong. We correct stories openly.
Readers' picks show on the front page.
Follow @AncientHeraldHQ on X for the day's finds as they come in.
Comments
House rules. Stay on the subject, be kind to other readers and say where your facts come from. No advertising and no personal attacks. The editor reads along and hides comments that break these rules.
Loading comments…